- Email: [email protected]
Fatal intersection crashes in Norway: Patterns in contributing factors and data collection challenges
Accident Analysis and Prevention 45 (2012) 782–791
Contents lists available at SciVerse ScienceDirect
Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap
Fatal intersection crashes in Norway: Patterns in contributing factors and data collection challenges Mikael Ljung Aust a,∗ , Helen Fagerlind b , Fridulv Sagberg a a b
Vehicle Safety Division, Department of Applied Mechanics, Chalmers University of Technology, SE-412 96 Göteborg, Sweden Institute of Transport Economics, Gaustadalléen 21, NO-0349 Oslo, Norway
a r t i c l e
i n f o
Article history: Received 6 July 2011 Received in revised form 20 October 2011 Accepted 2 November 2011 Keywords: Intersection crashes Causation analysis Contributing factors Driver behaviour analysis Indepth data collection
a b s t r a c t Fatal motor vehicle intersection crashes occurring in Norway in the years 2005–2007 were analyzed to identify causation patterns among their underlying contributing factors, and also to assess if the data collection and documentation procedures used by the Norwegian in-depth investigation teams produces the information necessary to do causation pattern analysis. 28 fatal accidents were analyzed. Causation charts of contributing factors were first coded for each driver in each crash using the Driving Reliability and Error Analysis Method (DREAM). Next, the charts were aggregated based on a combination of conflict types and whether the driver was going straight or turning. Analysis results indicate that drivers who were performing a turning maneuver in these crashes faced perception difficulties and unexpected behavior from the primary conflict vehicle, while at the same time trying to negotiate a demanding traffic situation. Drivers who were going straight on the other hand had less perception difficulties but largely expect any turning drivers to yield, which led to either slow reaction or no reaction at all. In terms of common contributing factors, those often pointed to in literature as contributing to fatal crashes, e.g. high speed, drugs and/or alcohol and inadequate driver training, contributed in 12 of 28 accidents. This confirms their prevalence, but also shows that most drivers end up in these situations due to combinations of less auspicious contributing factors. In terms of data collection and documentation, there was an asymmetry in terms of reported obstructions to view due to signposts and vegetation. These were frequently reported as contributing for turning drivers, but rarely reported as contributing for their counterparts in the same crashes. This probably reflects an involuntary focus of the analyst on identifying contributing factors for the driver held legally liable, while less attention is paid to the driver judged not at fault. Since who to blame often is irrelevant from a countermeasure development point of view, this underlying investigator approach needs to be addressed to avoid future bias in crash investigation reports. © 2011 Elsevier Ltd. All rights reserved.
1. Introduction Active safety functions, or Advanced Driving Assistance Systems (ADAS), are receiving an increasingly prominent role in traffic safety. The goal of these functions is to prevent crashes from occurring and/or to reduce crash severity, by either alerting the driver to hazards or by taking over the driving task to some extent, e.g. by autonomous braking or steering in emergency situations. Examples of ADAS available on the market include Forward Collision Warning (FCW) and Lane Departure Warning (LDW). To develop relevant ADAS, and to evaluate the extent to which they prevent and/or mitigate crashes, it is essential to be able to characterize the sequence of events which leads to crashes, in a way that includes information on the contributing factors that underlie
∗ Corresponding author. Tel.: +46 31 772 3677. E-mail address: [email protected] (M. Ljung Aust). 0001-4575/$ – see front matter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.aap.2011.11.001
the crashes (Najm et al., 1995; Najm and Smith, 2002). One source from which to derive such pre-crash characterization information is in-depth crash investigation data, or microscopic data (OECD, 1988). According to Midtland et al. (1995) and Larsen (2004), such qualitative in-depth crash information is the best source available for identifying interactions between crash contributory factors, i.e. for defining crash causation mechanisms. In terms of available in-depth crash data (at least in the Nordic countries) most of the available information comes from fatal crashes investigated by national road authorities. This is a natural consequence of the injury reducing strategy these authorities have employed, i.e. by focusing their investigation resources on the crashes with most severe outcomes and looking for ways in which to prevent those outcomes, the number of road deaths and severe injuries has successfully been reduced over the years. However, in terms of developing and evaluating ADAS and other countermeasures, a problem with fatal crashes is that one or more driver/occupant narratives always will be missing. Thus, while
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
783
Fig. 1. Distribution of Norwegian police reported vehicle crashes (i.e. including PTW’s but excluding VRU’s) based on frequency count, fatality count and injury count in the years 2005–2007 (Statistics Norway, 2011).
in-depth investigation of fatal crashes has proven very valuable for improvement of road and vehicle design, it is not clear to what extent current in-depth study procedures employ the data collection protocols and has access to the information required to produce in-depth studies that are helpful for ADAS design and evaluation. There were two aims for the present study. The first was to try to identify underlying contributing factors and causation patterns in a set of fatal crashes, based on reports from the road authorities’ in-depth investigation teams. The second aim was to assess the extent to which these factors are relevant for ADAS design and development, i.e. whether the reports produces contain the information necessary to perform ADAS relevant causation pattern analysis. To limit the study scope, the analysis was performed on intersection crashes only. Intersection crashes were selected because it is a crash type for which few, if any, ADAS have been deployed due to technical and other challenges involved, and a more precise problem description would be very helpful to their development (Bärgman and Smith, 2009). Furthermore, the study scope was limited to motor-vehicle crashes, i.e. crashes with vulnerable road user crashes (VRU’s, e.g. pedestrians and cyclists) were excluded, while crashes with Powered Two Wheelers (PTW), such as motorcycles and mopeds, were retained. VRU’s were excluded since they move and behave differently than motor vehicles; they use different parts of the road space, the kinematics are different, threat perception is made on different grounds, etc. A causation analysis which combines motor vehicle drivers with VRU’s therefore risks “hybrid” findings that may not be representative of either group. 2. Methodology The in-depth investigation reports used in the study come from Norway. In Norway all fatal crashes are analyzed in-depth by multidisciplinary crash investigation teams organised by the Norwegian Public Roads Administration (NPRA). The NPRA teams collect a mixture of data from on-the-scene investigations (i.e. right after the crash when the vehicles still are in place, mainly based on what the police observes) and/or on-the-site investigations (members of the team visit the crash site after the scene has been cleared to
Table 1 Vehicle sizes involved in the 28 fatal intersection crashes. Vehicle size Small Medium Large
Motorcycle Passenger car or similar Truck
Total
All
Turning
Going Straight
17 30 10
0 26 2
17 4 8
57
28
29
collect additional data). Also, records from later police interviews with the involved parties are collected. When all data is compiled, they produce an investigation report for each crash. The data for this study includes all 28 fatal motor vehicle intersection crashes that occurred in Norway in the years 2005–2007. In this sense, the sample is nationally representative of the crash type. However, 28 is a relatively low number given that there were 559 fatal crashes in total during that time period. Fatal crashes can thus be said to largely occur outside intersections, something which is further illustrated in Fig. 1 (data from Statistics Norway, 2011). This figure shows the relative frequency distributions over the major crash types for motor vehicle crashes and their associated fatalities and injuries. While crash and injury frequencies show similar distributions among crash types, fatalities have a rather different distribution. The implications of this in terms of result generalizability are further discussed Section 4. The data available for each crash varied in scope; for some crashes only the final report was available (usually a 5–10 page document), whereas for other crashes various protocols filled out during the investigation were also available. All crashes involved two vehicles except one, which involved three. There were three car-to-car crashes, whereasthe other 25 occurred between vehicles of pronounced size difference (i.e. PTW-to-car, car-to-truck). In Table 1, an overview of the involved vehicles is shown, along with information on to what extent vehicles in each size group was turning or going straight through the intersection. 2.1. Analysis procedure Prior to the analysis of contributing factors, the 28 cases were sorted following an intersection crash typology developed in a
784
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
LTAP/OD
LTAP/LD
SCP
Le Turn Across Path/ Opposite Direcon
Le Turn Across Path/ Lateral Direcon
Straight Crossing Paths
LTIP
RTIP
Le Turn Into Path
Right Turn Into Path
Fig. 2. Intersection crash types as defined by Najm et al. (2001).
Table 2 Distribution of crash types in analyzed material according to the typology from Najm et al. (2001). Fatal intersection accident conflict patterns LTAP-OD
LTAP-LD
SCP
RTIP
LTIP
Total
13
10
1
2
2
28
statistical study to describe crossing path crashes in the US (Najm et al., 2001). The typology sorts crashes based on actual and intended vehicle trajectories prior to the crash (Fig. 2). As an example, the typical scenario for a Left-Turn-Across-Path/Opposite Direction (LTAP/OD) crash is a left turning vehicle cutting across the path of another vehicle coming from the opposite direction, and which intends to cross the intersection on a straight path. This typology was used because it divided the US intersection crash population into subgroups that were clearly differentiated in terms of the typical circumstances under which they occur (Najm et al., 2001). Thus it might be expected that they could be differentiated in terms of contributing factors and causation patterns as well. Moreover, using vehicle trajectories as a typology basis provides a very natural frame of reference and principles for separation for the technical side of ADAS development, as vehicles on turning trajectories likely will have different sensor and Human Machine Interaction (HMI) requirements than vehicles going straight through an intersection. The distribution of the 28 crashes according to the typology is presented in Table 2. The most frequent conflict patterns are LTAPOD and LTAP-LD, which also were among the top three in Najm et al. (2001). However, the most common pattern found by Najm et al. (SCP crashes) is nearly absent from this sample. This is somewhat surprising, as there is no obvious reason why fatal crashes should be differently distributed in these conflict patterns compared to crashes in general. Unfortunately the Norwegian crash
typology can separate vehicles based on whether they came from the same or different roads but not based on whether they intended to go straight or turn. This means that LTAP-LD crashes cannot be separated from SCP crashes, which makes a general US–Norway comparison impossible. When the cases had been sorted, details on crash contributing factors for each driver in each in-depth case were coded using the Driving Reliability and Error Analysis Method (DREAM 3.0) (Wallén Warner et al., 2008). DREAM is an adaptation to the traffic safety domain of the Cognitive Reliability and Error Analysis Method (CREAM) (Hollnagel, 1998). The first version (Ljung, 2002) was developedin the FICA project (Ljung et al., 2007). It has since then been iteratively refined. The version used here was developed and applied in the European project SafetyNET (SAFETYNET, 2008), where approximately 1000 in-depth investigated crashes were DREAM coded. DREAM has two purposes: to support systematical classification of accident causation information, and to facilitate aggregation of that information into patterns of contributing factors (i.e. causation charts).1 DREAM includes three main elements: an accident model, a classification scheme and a method for applying the scheme to investigated events. The accident model serves as a framework and theoretical anchor for the classification scheme and method. It uses the human–technology–organisation (HTO) triad as a reference – represented by the driver (human), the vehicle and traffic environment (technology) and the organisation. The driving task is characterized through the Extended Control Model (ECOM; Hollnagel and Woods, 2005), which defines driving as a control task which includes working towards multiple parallel
1 It is worth pointing out that DREAM (like most other tools for accident analysis) is an organiser of explanations – not a provider. For any classification code in DREAM to be applicable, it must be supported by empirical information. If no information exists, there is nothing to classify.
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791 Table 3 Phenotypes in DREAM 3.0. Phenotypes
Specific phenotypes
Timing Speed Distance Direction Force Object
Too early action; too late action; no action Too high speed; too low speed Too short distance Wrong direction Surplus force; insufficient force Adjacent object
goals on different time scales. Accidents happen when an unsuccessful interplay between driver, vehicle and traffic environment leads to loss of control in one or more control processes, and that control cannot be regained given available time and/or resources. The classification scheme comprises a number of observable effects in the form of human actions and system events called phenotypes. Phenotypes are expressed in the general dimensions of time, space and energy, and consist of the following (Table 3). It also contains a number of possible contributing factors which may have brought about these observable effects. These are called genotypes, and are organised according to the driver–vehicle/traffic environment–organisation triad mentioned above. The driver category contains cognitive function genotypes related to problems with observation, interpretation and planning (in accordance with ECOM). It also includes more general person related factors that may degrade driving performance and hence contribute to an accident (e.g. fatigue). The other categories contain vehicle, traffic environment and organisation related contributing factors. The genotypes include contributing factors both at the sharp end (close in time/space to the crash) as well as at the blunt end (more distant in time/space, yet important for the development of events). In DREAM version 3.0 which was used here, genotypes are divided into 16 main categories, each belonging to one of four main groups: Driver, Vehicle and Traffic environment and Organisation (see Table 4). The phenotype/genotype distinction is adapted from biology, and refers to the fact that while for example all humans look more or less different (each phenotype is unique), we have share an identical genotype (a pool of genes), and differences between individuals depend on which combination of genes that was active in the shaping of that phenotype. Vehicle crashes can be viewed analogously, i.e. while each crash is unique, there is a common set of possible contributing factors that can be used to analyze (most) crashes. DREAM also includes a link system which specifies possible interactions between contributing factors. When case information on causation is coded into a chart, the link system ensures that the description of how one contributing factor leads to another is not arbitrary. The link system basically limits the range of possible factor interactions to those currently supported by scientific knowledge, thus restricting and guiding the coding of causation information. The inherent structure in the link system also makes it possible to aggregate causation information from multiple case studies in a structured, and principally semi-automated fashion, reducing the number of subjective judgements necessary to identify a pattern of contributing factors for a group of crashes. Naturally, the link system can be updated as new knowledge is gained. For a discussion of how to create and interpret aggregated causation charts using DREAM, see (Sandin and Ljung, 2007; Sandin, 2008). Lastly, the method for how to do the actual classification contains a detailed step-by-step instruction, as well as several stop rules. The latter are well defined conditions that determine when the analysis should come to an end. These are necessary, as the classification scheme represents a network rather than a hierarchy, and deciding when to stop the analysis can become arbitrary in the absence of such rules.
785
An assumption underlying the coding was that each driver has his/her own reasons for failing to adapt to the driving situation. Contributing factors were therefore coded separately for each involved driver, resulting in one causation chart per involved driver. Next, the individual charts for each crash type were aggregated based on which conflict trajectory (Fig. 2) the driver was on. Thus, two aggregate causation charts were created per crash type; one for turning drivers and one for drivers going straight. For the SCP crash, in which both drivers were going straight, only one aggregated chart was created. All coding was performed by the first author of this paper. 3. Results In the figures below, the total number of times a contributing factor occurs is represented by the number in brackets within each box. Note that the way DREAM is structured allows attribution of, for example, multiple missed observations to a single driver, depending on how many factors contributed to the missed observation. The occurrence frequency for some contributing factors might therefore exceed the number of aggregated causation charts. For visual guidance when looking for patterns, the factor frequency numbers are also indicated through box border thickness. For links between boxes, the number of times a link occurs is not written out, but indirectly represented through the thickness of the connecting arrows. Note also that to make the aggregated charts readable, the detailed information from the individual charts which motivates the choice of each phenotype and genotype has been removed. However, in the analysis, the detailed information from the individual charts was still used. The full analysis of each crash can be found in the final project report where the analysis was carried out (Akhtar et al., 2010). 3.1. LTAP/OD crashes The first analyzed crash type is Left Turn Across Path/Opposite direction (LTAP/OD) crashes. As described above, drivers going straight and turning drivers were put in separate groups prior to the aggregation. Below, the aggregate causation patterns for each of these groups are presented (Figs. 2 and 3), and the details of the most common contributing factors are described. 3.1.1. Left-Turn-Across-Path (LTAP) drivers While there were 13 LTAP-OD crashes in the sample, there are 14 phenotypes (and hence 14 LTAP drivers) described in Fig. 3. This is because one of the crashes involved three vehicles, and two of these were turning. A majority of the LTAP drivers involved in LTAP-OD crashes were coded with the phenotype Timing: too early (13 of 14). This indicates that they began to turn through the intersection before it was appropriate to do so. Here, “before it was appropriate” refers to the fact that these turning drivers were expected to yield to any vehicle on a straight crossing path before turning, which they in these cases did not do. Two of those too early turns are preceded by Misjudgment of time gaps, i.e. the turning driver overestimates the time available for completing the turn before the oncoming vehicle reaches the intersection. The other 11 are coded with Misjudgment of situation as the main contributing factor. Contributing to those misjudgments are 8 instances of Missed observation, meaning that the driver did not see the other vehicle at the time when s/he decided to initiate the turn. A closer reading of the individual charts shows that the conflict vehicle in 7 of those 8 instances was a PTW, which the drivers were unable to perceive at “checkpoint time”, i.e. when they decided it was okay to make the turn. Reasons for not perceiving the vehicles
786
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
Table 4 Genotypes in DREAM 3.0. Genotypes (B–Q) Human (B–F)
B: Observation Missed observation (B1) Late observation (B2) False observation (B3)
Technology (G–M) Vehicle (G–I)
Traffic environment (J–M)
G: Temporary HMI problems Temporary illumination problems (G1) Temporary sound problems (G2) Temporary access limitations (G4)
J: Weather conditions Reduced visibility (J1)
N: Organisation Time pressure (N1)
Strong side winds (J2)
Incorrect ITS-information (G5)
K: Obstruction of view due to object Temporary obstruction of view (K1) Permanent obstruction of view (K2)
Irregular working hours (N2) Heavy physical activity before drive (N3) Inadequate training (N4)
C: Interpretation Misjudgement of time gaps (C1)
H: Permanent HMI problems
Misjudgement of situation (C2)
Permanent illumination problems (H1) Permanent sound problems (H2) Permanent sight obstruction (H3)
D: Planning Priority error (D1)
E: Temporary personal factors
Organisation (N–Q)
I: Vehicle equipment failure
L: State of road Insufficient guidance (L1) Reduced friction (L2) Road surface degradation (L3)
Equipment failure (I1)
Object on road (L4)
Fear (E1)
Inadequate road geometry (L5)
Inattention (E2) Fatigue (E3) Under the influence of substances (E4) Excitement seeking (E5) Sudden functional impairment (E6) Psychological stress (E7)
M: Communication Inadequate transmission from other road users (M1) Inadequate transmission from road environment (M2)
O: Maintenance Inadequate vehicle maintenance (O1) Inadequate road maintenance (O2) P: Vehicle design Inadequate design of driver environment (P1) Inadequate design of communication devices (P2) Inadequate construction of vehicle parts and/or structures (P3) Unpredictable system characteristics (P4) Q: Road design Inadequate information design (Q1) Inadequate road design (Q2)
F: Permanent personal factors Permanent functional impairment (F1) Expectance of certain behaviours (F2) Expectance of stable road environment (F3) Habitually stretching rules and recommendations (F4) Overestimation of skills (F5) Insufficient skills/knowledge (F6)
going straight include Reduced visibility (here: MC driver wearing dark clothing against dark background), Inadequate transmission from other road users (here: sub-standard MC lighting) and various obstructions to view. The latter include obstructions inside the vehicle (Temporary sight obstruction, here items on the dashboard in one case and a dirty windshield in the other), and obstructions outside the vehicle. For outside distractions, Temporary obstruction of view indicates blocked lines of sight due to other vehicles in the traffic environment at critical decision moments, while the objects coded with Permanent obstruction of view primarily refers to vegetation. There is however also one instance of a traffic sign blocking the line of sight, and one instance where the road lighting design actually reduced object visibility in rain. Apart from difficulties in perceiving the road user going straight, the turning drivers also have 8 instances of Inattention coded as contributors to Misjudgment of situation. These instances of inattention are largely attributed to Driving-related distractors outside vehicle (6 instances), which means that the drivers were focused on negotiating some other relevant task in the traffic environment. In four of the six instances, this involved tracking a third vehicle also about to enter the intersection. The fifth instance involved tracking pedestrians who are crossing the road somewhat randomly (outside a school), and the last involved maneuvering a large truck and trailer combination through a (for that vehicle combination) tight turn.
Another frequent contributor to Missed observation is Expectance of certain behaviours, with four instances. Three of these refers to drivers who expected other vehicles to keep approximately to posted speed limits and who therefore did not look sufficiently far down the road to discover a MC travelling at very high speed.
3.1.2. OD drivers There were 13 drivers on a straight crossing path in the LTAP-OD crashes, and the aggregated causation chart for these is shown in Fig. 3. The phenotype coding for these drivers is more diverse than for the turning drivers. 5 are coded as Timing: no action, indicating that they did not perform any type of steering and/or braking prior to the crash. The 4 coded as Timing: too late have to some extent started a corrective action, but to late to avoid the crash. 3 drivers are coded as Speed: too high. Two involve MCs travelling at substantially higher speeds than legally permitted. The third case involves a car which negotiates a curve with very limited visibility at slightly above legal speed, which leaves the driver with insufficient time to respond to a vehicle standing still at an intersection just after the curve. The last driver is coded as Force: insufficient force, which here means that if the driver had used the normal braking capacity of the MC he was driving the crash most likely would have been avoided (Fig. 4).
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
787
Fig. 3. Aggregated causation chart for the 14 drivers involved in LTAP/OD crashes who were on a LTAP trajectory.
The most common contributing factor for OD drivers is Misjudgment of situation (12 instances), 10 of which are preceded by Expectance of certain behaviours. Here this indicates that the drivers going straight were expecting any turning vehicles to yield, i.e. the drivers have most likely seen the other vehicle, but expects it to stop/wait for them. There are also three drivers for whom the chain Inadequate training → Insufficient skills/knowledge leads to Misjudgment of situation. Two of these lacked a drivers license for the MC they were driving. These two were also coded as Under the influence of substances, as well as with the chain Excitement seeking → Priority error. The third driver was on a high performance MC, the full brake capacity of which he did not master. 3.2. LTAP/LD crashes The second analyzed crash type is Left Turn Across Path/Lateral Direction (LTAP/LD) crashes. As described above, drivers going straight and turning drivers were put in separate groups prior to the aggregation. Below, the aggregate causation patterns for each of these groups are presented (Figs. 6 and 7), and the details of the most common contributing factors are described. 3.2.1. Left-Turn-Across-Path (LTAP) drivers The aggregated causation pattern for the 10 turning drivers in LTAP-LD crashes is shown below in Fig. 5. As can be seen in the chart, the turning drivers in LTAP-LD crashes are, just like the turning drivers in LTAP-OD crashes, almost exclusively phenotype coded as Timing: too early (9 of 10). As for LTAP-OD crashes, this indicates that they began to turn through the intersection before it was appropriate to do so. The immediate antecedent for all 10 crashes is Misjudgment of situation. Contributing to that misjudgment are 7 instances of Missed observation, to which in turn a combination of Permanent obstruction to view and Inattention contributes. As above, the instances of Permanent obstruction to view mainly refer to
vegetation elements in the intersection (6 of 7 instances), but traffic environment elements also play a part (2 signposts and one case of inadequate intersection layout). The 6 instances of Inattention are mainly attributed to distractors, the most common being the 4 instances of Driving related distractors outside vehicle, e.g. other vehicles which must be considered when negotiation of the traffic situation. There is also one instance of a Driving related distractor inside vehicle and one instance where Inattention is attributed to the driver being drunk (Alcohol). In two of the three cases where Insufficient skills/knowledge has been coded, this refers to young (18 years) and relatively inexperienced drivers who have had their driving licenses for a short time. The third case involves a driver who is negotiating the manual shift of a rental car, while being used to automatic gear in his own vehicle. 3.2.2. OD drivers The aggregated causation pattern for the 10 drivers going straight in LTAP-LD crashes is shown below in Fig. 6. Five of these are coded as Timing: no action, which means that they did not perform any type of steering and/or braking prior to the crash. The 3 coded as Timing: too late have initiated corrective actions, but too late to avoid the crash. One driver is coded as Speed: too high, indicating a travel speed substantially over the speed limit. In terms of contributing factors, the most common contributor is Misjudgment of situation. As for OD drivers in LTAP-OD crashes, a majority of these are in turn attributed to Expectance of certain behaviours, i.e. the drivers expects non-priority vehicles to yield. There are also 2 instances of Missed observation and 2 instances of Late observation contributing to Misjudgment of situation. The instances of Missed observation are in turnattributed to one instance of Inadequate transmission from other road users and one instance of Permanent obstruction to view. There is one instance of Misjudgment of time gaps, where an MC driver overestimated the time available for passing through the intersection before the turning vehicle would be blocking the MC’s
788
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
Fig. 4. Aggregated causation chart for the 13 drivers involved in LTAP/OD crashes who were on an OD trajectory.
travel path. This driver is also coded as Excitement seeking → Priority error and Alcohol → Under the influence of substances and Speed: too high.
as trying to beat a third vehicle (coming from his right) to the intersection by driving fast (the conflict vehicle came from his left). 3.4. Right-Turn-Into-Path (RTIP) crashes
3.3. SCP crashes There was just one SCP crash in the material. Here, both drivers were coded with Timing: no action, indicating that neither made any attempt to stop or yield as they approached the intersection. Both drivers had difficulties seeing the other vehicle (Late observation and Missed observation respectively), due to a mutual Permanent obstruction of view. One driver is also coded as Excitement seeking → Priority error in combination with Driving-related distractors outside vehicle. Here, this refers to the fact that the driver was reported
There were two RTIP crashes in the sample. Of the drivers turning right, one is coded with Timing: too early, indicating that he commenced the turn before the intersection was clear, while the other is coded as Timing: no action, indicating that he drove into the intersection without yielding to the vehicle going straight. Contributing factors are one instance of Misjudgement of situation due to Missed observation and Expectance of certain behaviourism, which refers to one of the drivers who was not looking for vehicles travelling at speeds much higher than posted speed limit, and who
Fig. 5. Aggregated causation chart for the 10 drivers involved in LTAP/LD crashes who were on a LTAP trajectory.
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
789
Fig. 6. Aggregated causation chart for the 10 drivers involved in LTAP/LD crashes who were on a LD trajectory.
therefore did not look sufficiently far down the road to see the MC approaching at high speed. Of the two drivers going straight one was speeding on a motorcycle but expected the turning driver to yield. In the other case, the right turning car just pulled out in front of a semitrailer, and the trailer driver did not have time to react at all. 3.5. Left-Turn-Into-Path (LTIP) crashes There were two LTIP crashes in the sample. Both turning drivers commenced the turn prematurely, (Timing: too early) due to Missed observation (s/he did not observe an oncoming MC). In both cases, the approaching MC was travelling much faster than the speed limit and was difficult to perceive in terms of optical visibility. The MC drivers on a straight path in these crashes both were driving much faster than legally permitted (both coded as Excitement seeking→Priority error). One of the drivers was involved in what can be referred to as a peer pressure situation, i.e. wanting to show off to a friend (Psychological stress), and overestimating his own driving skills (Overestimation of skills). 4. Discussion There were two aims in the present study. The first was to try to identify underlying contributing factors and causation patterns relevant for ADAS design and development in intersection crashes, based on reports from a set of in-depth investigated fatal crashes. The second aim was to evaluate whether the data collection and documentation procedures used to perform the in-depth investigations yield the information necessary to perform this type of causation pattern analysis. In terms of the first objective, it is clear that there is a lot of relevant information available in these in-depth fatal crash reports. Also, separating drivers based on trajectory following the intersection conflict typology developed by Najm et al. (2001) is useful in many aspects. In the analysis performed here however, it seems warranted to aggregate also at a higher level, based on whether drivers are going straight or turning. In other words, when looking at the causation charts, it seems like turning drivers have contributing factors in common across the Najm conflict types, as do drivers going straight, and there seems to be a clear difference between these two groups. The patterns in each respective group, and what the implications of those are for ADAS design, will be discussed below. After that, the discussion will come back to the data quality issue, i.e. the relationship between data collection and documentation procedures and analysis outcome.
4.1. Common factors for turning drivers Turning drivers are almost uniformly coded with timing: too early, indicating that they initiate the turn before the crossing vehicle has passed. In terms of ADAS development, this clearly indicates a need for turning decision support, i.e. smart ways of helping the driver understand when it is safe to initiate a turn and when it is not. Moreover, the misjudgements of situation underlying these too early turns seems to be largely driven by two types of problems. One is that the drivers do not see the conflict vehicle at the time of making the decision to turn, due to physical obstructions to view. In terms of ADAS design, this means that regular, vehicle based, sensors such as cameras and radar will not be sufficient to solve the problem, as their lines of sight would be hindered as well. Instead, vehicle-to-vehicle communication technology probably needs to be considered. Second, a large number of the turning drivers are focused on some other part of the traffic situation than the oncoming vehicle (Driving-related distractors outside the vehicle). This means than an information/warning ADAS not only needs to inform the driver about a problem, it needs to do so in a way that “breaks through” the drivers’ current focus, i.e. the ADAS has to redirect attention rather than just get attention. Consequently, evaluation of an information/warning ADAS for turning drivers at intersections needs to be performed in a situation where the test drivers really are engaged in some concurrent, high priority, driving related task (like tracking another vehicle about to enter the intersection). If this condition is not met, the ADAS ability to redirect the drivers’ attention will not be properly assessed. This is further underlined by the fact that turning drivers are also frequently coded with Expectance of certain behaviours, referring to turning drivers not adjusting their regular intersection scanning pattern to accommodate vehicles travelling at speeds well above the speed limit. In other words, the behaviour of the conflict vehicle lie outside what normally can be expected in many of the cases analyzed here. This circumstance also needs replicating in an evaluation situation. 4.2. Common factors for drivers going straight Drivers going straight show a high frequency of planning failures due to Expectance of certain behaviours. This likely reflects the fact that drivers going straight often have the right of way and therefore expect turning vehicles to yield. Adding to this interpretation is the fact that drivers going straight are much less frequently coded with Missed observation and Obstructions to view than turning drivers. Given that the driver going straight has the right of way, the task of
790
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
identifying and responding to a conflict vehicle in an intersection rests with the turning driver. Sight limitations are thus more debilitating for the turning driver’s performance, and hence more likely to be reported as a contributing factor for them than for drivers going straight. In terms of ADAS development and evaluation, this means than an information/warning ADAS not only needs to inform the driver going straight about a problem, it needs to do so in a way that convinces a driver who believes s/he has the right of way to actually brake and/or swerve. This sets a very challenging target for the ADAS threat assessment. First, for a driver going straight to trust and rely on such an ADAS it has to be right most of the time, i.e. being perceived by the driver as making relevant and warranted calls rather than crying wolf. Second, the threat assessment needs to be carried out sufficiently well ahead in time to allow the driver to act on a warning. It is worth noting that the group of drivers going straight contains all 17 PTW drivers involved in the 28 fatal crashes, many of which were driving well above the speed limit. Also, when it comes to contributing factors like Excitement seeking and Alcohol/Drugs, as well as Overestimation of skills and Insufficient skills/knowledge, these are almost exclusively coded for drivers going straight. In terms of ADAS evaluation and development, this presents some interesting strategy choices. If one wants to develop an ADAS that would work for all drivers going straight, this means that the test drivers at evaluation somehow need to represent the conditions which the coded factors indicate (excitement seeking, alcohol, etc.) for the assessment to be correct. This presents interesting challenges for test subject preparation and evaluation. It is also challenging for the ADAS technical capacity, considering the actual travel speeds that needs to be handled. If one on the other decides to develop an ADAS for the “normal” driver, i.e. not speeding, untrained, etc., then the number of drivers which the ADAS actually can help may be quite limited. 4.3. Data quality A general observation is that there are in general fewer contributing factors coded for drivers going straight than for turning drivers. In other words, despite the investigators being trained observers tasked with taking an objective view of the data, they seem more likely to provide deeper and fuller explanations of why the turning driver gets into trouble, compared to the driver going straight. While it might be that there simply is less relevant information to code for drivers going straight, a more likely reason for this focus on the turning drivers (albeit involuntary) is connected to the fact that the turning driver usually is the one held legally liable for the crash. Since the driver going straight normally has the right of way it is easy to conclude that it was not his/her fault, and investigation might therefore come to focus on explaining why the driver at legal fault ended up in the crash. An example which illustrates this asymmetry is the number of reported obstructions to view due to signposts and vegetation. Since these are part of the traffic environment, i.e. the infrastructure, one would assume that any blockage in lines of sight is reciprocal, i.e. if driver A cannot see driver B, the reverse should also be true. However, while external obstructions to view are frequently reported as contributing for turning drivers, they are rarely reported as contributing for their counterparts in the same crash. This investigator bias towards focusing on the party held liable presents a challenge for ADAS development. It is true from a physics standpoint that if two vehicles are on a colliding path, it will almost always be easier for one of the vehicles to perform the avoidance maneuver necessary to prevent a collision. However, it is far from clear that this vehicle necessarily belongs to the driver at fault; it is
a kinematic relationship between masses rather than a moral relationship between operators. In other words, who to blame is usually irrelevant from an ADAS development point of view. Underreporting of contributing factors for one of the parties involved based on reasoning about guilt thus hinders rather than helps countermeasure development. This underlying investigator mind-set, if it can be called that, would need addressing to avoid future bias in reported information. The amount of available data in this study varied somewhat between crashes; for some there was just the final report, whereas for other accidents various protocols filled out by the investigators during the investigation were available. The latter extra information several times proved to be very valuable for the analysis, not in the sense that it brought new information to light, but in order to rule possible contributing factors out. While the final reports overall are very good at compiling the relevant information from the other documents, they describe inclusions rather than exclusions. In other words, the reasons for why certain factors are thought to contribute are included, but the reasons for excluding other possible factors are left out. Here, the extra information in the other protocols could at times be used to discard certain possible contributing factors, the established absence of which certainly made a difference in terms of how the accident causation process was reconstructed. An interesting fact is that contributing factors often are pointed to as main contributors to fatal crashes, e.g. high speed, drugs/alcohol and inadequate driver training, played a role in 12 of 28 accidents. Two implications follow from this. First, if one were to evaluate an ADAS addressing intersection crashes, finding and preparing a relevant group of test persons would be a challenge, at least if the group is meant to be representative of the drivers involved in the crashes investigated here. For example, using intoxicated test persons has numerous ethical issues associated with it. Second, while this confirms the prevalence of these factors, it also indicates that most drivers end up in these situations due to combinations of less auspicious contributing factors. In this study, a number of the possible contributing factors that the DREAM methodology contains were never applied to any of the 28 crashes. As DREAM has been put through validation work and corroboration with other researchers’ findings on possible crash causes, there is reason to further investigate why many of the contributing factors available in DREAM never were applied. There are at least three possible explanations for this. One is that the information necessary for coding those genotypes usually comes from detailed driver interviews. As this is inherently impossible for at least one subject in a fatal crash, there is a risk that the analysis becomes systematically biased. However, unless contributing factors for the fatally wounded systematically differ from the factors for those who survived, the survivors should still provide information related to these factors if present. Another possible explanation is that the Norwegianin-depth investigation teams rarely conduct their own driver interviews. Rather they rely on information from the police interviews. Here it can be hypothesised that crash survivors are not always completely forthcoming when describing crash circumstances to the police. A third possibility is that the full set of contributing factors in DREAM might not be necessary for analysis of fatal crashes, i.e. the factors not coded are simply not relevant contributors in crashes with fatal outcomes. Which of these explanations are the best is for future studies. As all coding was performed by the author, the question of whether another coder would have come to different conclusions naturally arises. While assessing this topic was outside the scope of this study, a previous interrater reliability study showed a relatively high interrater agreement for coders applying DREAM to a set of sample cases (Wallén Warner and Sandin, 2010). Thus, while some variation must be expected if another coder would code the
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
same set of cases, the variability could be expected to stay within reasonable limits. This is however also a topic for future study. A final topic for discussion is the representativity of the contributing factors found for the 28 crashes studied here. On one hand, since the data set contains all crashes of the selected type that occurred in the selected years, the data set is clearly nationally representative. On the other hand, as Fig. 1 shows, fatalities have a different distribution across crash types than crashes and injuries in general. To extend the conclusions on contributing factors drawn for the data set here to crashes in general, one thus would need to explain the distribution difference in a way that does not involve the causation analysis. One possibility is to argue that the distribution difference reflects the energies involved rather than what brought the crash about. If so, Fig. 1 could be interpreted as showing that impact speeds are highest in head-on collisions and lowest in rear-end crashes, with single vehicle crashes and intersection crashes in between. However, while this explanation seems plausible, it does not a priori rule out the possibility that fatal crashes have at least somewhat different contributing factors than crashes in general. For example, if the numbers on high speed, drugs/alcohol and inadequate driver training above were representative of all crashes, then they would play a role in 4 of 10 intersection crashes, which seems quite high. Finding a way of deciding between these explanations is clearly a topic for future study. 5. Conclusions There were two aims in the present study. The first was to try to identify underlying contributing factors and causation patterns relevant for ADAS design and development in intersection crashes, based on reports from a set of in-depth investigated fatal crashes. The second aim was to evaluate whether the data collection and documentation procedures used to perform the in-depth investigations yield the information necessary to perform this type of causation pattern analysis. The study shows that it is indeed possible to identify such causation patterns in this type of data. This study indicates that turning drivers to a large extent are faced with perception difficulties and unexpected behaviours in relation to the conflict vehicle, while at the same time trying to negotiate a demanding traffic situation. Drivers going straight on the other hand have less perception difficulties. Instead, their main problem is that they largely expect turning drivers to yield. When this assumption is violated, they are either slow to react or do not react at all. These findings are very informative when it comes to development and evaluation of ADAS intended to address intersection crashes, as they present direction and place constraints on how the ADAS needs to interact with the driver to be successful. Furthermore, contributing factors often pointed to in literature, e.g. high speed, drugs and/or alcohol and inadequate driver training, played a role in 12 of 28 accidents. While this confirms their prevalence, it also indicates that most drivers end up in these situations due to combinations of less auspicious contributing factors. In terms of data collection and documentation, the analysis shows that there seems to be an (involuntary) analyst focus on
791
identifying contributing factors for the driver legally held liable, which in this crash type often is the turning driver, while less attention is paid to the driver not at fault (the driver going straight). However, since who to blame often is irrelevant from a countermeasure development point of view, this underlying mind-set needs addressing to avoid future bias in the reported information. Acknowledgements This study draws on accident investigations performed by the Norwegian Public Roads Administration, and was funded by the Research Council of Norway, under the RISIT programme (Risk and Safety in Transport). References Akhtar, J., Ljung Aust, M., Eriksson, R.J., Fagerlind, H., Høye, A., Phillips, R., Sagberg, F., 2010. Factors Contributing to Road Fatalities: Analysis of In-depth Investigation Data from Passenger Car Intersection Crashes and From Collisions Between Bicycle and Motorized Vehicle. Oslo: Institute of Transport Economics, Oslo, Norway, TØI Report 1067. Bärgman, J., Smith, K., 2009. Ivss Intersections Accidents: Analysis and Prevention. Final Report. IVSS Programme, Borlänge, Sweden. Hollnagel, E., Woods, D.D., 2005. Joint cognitive systems: foundations of cognitive systems engineering. CRC Press, Boca Raton, FL. Hollnagel, E., 1998. Cognitive Reliability and Error Analysis Method (CREAM). Elsevier Science. Larsen, L., 2004. Methods of multidisciplinary in-depth analyses of road traffic accidents. Journal of Hazardous Materials 111, 115–122. Ljung, M., 2002. DREAM: Driving Reliability and Error Analysis Method. Master Thesis. Linköping University. Ljung, M., Fagerlind, H., Lövsund, P., Sandin, J., 2007. Accident investigations for active safety at chalmers – new demands require new methodologies. Vehicle System Dynamics 45 (10), 881–894. Midtland, K., Muskaug, R., Sagberg, F., Jørgensen, N.O.C., 1995. Evaluation of the In-depth Accident Investigations of the Swedish National Road Administration. Najm, W.G., Mironer, M., Koziol, J., Wang, J.S., Knipling, R.R., 1995. Synthesis Report: Examination of Target Vehicular Crashes and Potential Its Countermeasures. U.S. Department of Transportation, Volpe National Transportation Systems Center, Cambridge. Najm, W.G., Smith, J.D., 2002. Year Breakdown of light vehicle crashes by pre-crash scenarios as a basis for countermeasure development. In: Proceedings of the Proceedings of the International Conference on Applications of Advanced Technologies in Transportation Engineering, pp. 712–719. Najm, W.G., Smith, J.D., Smith, D.L., 2001. Analysis of Crossing Path Crashes U.S. Department of Transportation. John A.Volpe National Transportation Systems Center, Cambridge. OECD,1988. Road accidents: on-site investigations. In: Development. O.F.E.C.-O.A., Paris. SAFETYNET, 2008. Deliverable 5.8: In-depth Accident Causation Database and Analysis Report. European Road Safety Observatory. Sandin, J., 2008. Aggregating Case Studies of Vehicle Crashes by Means of Causation Charts – An Evaluation and Revision of the Driving Reliability and Error Analysis Method. Chalmers University of Technology. Sandin, J., Ljung, M., 2007. Understanding the causation of single-vehicle crashes: a methodology for in-depth on-scene multidisciplinary case studies. International Journal of Vehicle Safety 2 (3), 316–333. Statistics Norway, 2011. Tabell 03443: Personer drept eller skadd i veitrafikkulykker, etter trafikantgruppe og ulykkesgruppe (avslutta serie). Statistikkbanken område 10.12: Transport og kommunikasjon. Statistisk Sentralbyrå, Oslo. Wallén Warner, H., Ljung Aust, M., Björklund, G., Johansson, E., Sandin, J., 2008. Manual for DREAM 3.0, Driving Reliability and Error Analysis Method. Deliverable D5.6 of the EU FP6 Project SafetyNet. European Commission, Brussels. Wallén Warner, H., Sandin, J., 2010. The intercoder agreement when using the driving reliability and error analysis method in road traffic accident investigations. Safety Science 48 (5), 527–536.
Contents lists available at SciVerse ScienceDirect
Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap
Fatal intersection crashes in Norway: Patterns in contributing factors and data collection challenges Mikael Ljung Aust a,∗ , Helen Fagerlind b , Fridulv Sagberg a a b
Vehicle Safety Division, Department of Applied Mechanics, Chalmers University of Technology, SE-412 96 Göteborg, Sweden Institute of Transport Economics, Gaustadalléen 21, NO-0349 Oslo, Norway
a r t i c l e
i n f o
Article history: Received 6 July 2011 Received in revised form 20 October 2011 Accepted 2 November 2011 Keywords: Intersection crashes Causation analysis Contributing factors Driver behaviour analysis Indepth data collection
a b s t r a c t Fatal motor vehicle intersection crashes occurring in Norway in the years 2005–2007 were analyzed to identify causation patterns among their underlying contributing factors, and also to assess if the data collection and documentation procedures used by the Norwegian in-depth investigation teams produces the information necessary to do causation pattern analysis. 28 fatal accidents were analyzed. Causation charts of contributing factors were first coded for each driver in each crash using the Driving Reliability and Error Analysis Method (DREAM). Next, the charts were aggregated based on a combination of conflict types and whether the driver was going straight or turning. Analysis results indicate that drivers who were performing a turning maneuver in these crashes faced perception difficulties and unexpected behavior from the primary conflict vehicle, while at the same time trying to negotiate a demanding traffic situation. Drivers who were going straight on the other hand had less perception difficulties but largely expect any turning drivers to yield, which led to either slow reaction or no reaction at all. In terms of common contributing factors, those often pointed to in literature as contributing to fatal crashes, e.g. high speed, drugs and/or alcohol and inadequate driver training, contributed in 12 of 28 accidents. This confirms their prevalence, but also shows that most drivers end up in these situations due to combinations of less auspicious contributing factors. In terms of data collection and documentation, there was an asymmetry in terms of reported obstructions to view due to signposts and vegetation. These were frequently reported as contributing for turning drivers, but rarely reported as contributing for their counterparts in the same crashes. This probably reflects an involuntary focus of the analyst on identifying contributing factors for the driver held legally liable, while less attention is paid to the driver judged not at fault. Since who to blame often is irrelevant from a countermeasure development point of view, this underlying investigator approach needs to be addressed to avoid future bias in crash investigation reports. © 2011 Elsevier Ltd. All rights reserved.
1. Introduction Active safety functions, or Advanced Driving Assistance Systems (ADAS), are receiving an increasingly prominent role in traffic safety. The goal of these functions is to prevent crashes from occurring and/or to reduce crash severity, by either alerting the driver to hazards or by taking over the driving task to some extent, e.g. by autonomous braking or steering in emergency situations. Examples of ADAS available on the market include Forward Collision Warning (FCW) and Lane Departure Warning (LDW). To develop relevant ADAS, and to evaluate the extent to which they prevent and/or mitigate crashes, it is essential to be able to characterize the sequence of events which leads to crashes, in a way that includes information on the contributing factors that underlie
∗ Corresponding author. Tel.: +46 31 772 3677. E-mail address: [email protected] (M. Ljung Aust). 0001-4575/$ – see front matter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.aap.2011.11.001
the crashes (Najm et al., 1995; Najm and Smith, 2002). One source from which to derive such pre-crash characterization information is in-depth crash investigation data, or microscopic data (OECD, 1988). According to Midtland et al. (1995) and Larsen (2004), such qualitative in-depth crash information is the best source available for identifying interactions between crash contributory factors, i.e. for defining crash causation mechanisms. In terms of available in-depth crash data (at least in the Nordic countries) most of the available information comes from fatal crashes investigated by national road authorities. This is a natural consequence of the injury reducing strategy these authorities have employed, i.e. by focusing their investigation resources on the crashes with most severe outcomes and looking for ways in which to prevent those outcomes, the number of road deaths and severe injuries has successfully been reduced over the years. However, in terms of developing and evaluating ADAS and other countermeasures, a problem with fatal crashes is that one or more driver/occupant narratives always will be missing. Thus, while
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
783
Fig. 1. Distribution of Norwegian police reported vehicle crashes (i.e. including PTW’s but excluding VRU’s) based on frequency count, fatality count and injury count in the years 2005–2007 (Statistics Norway, 2011).
in-depth investigation of fatal crashes has proven very valuable for improvement of road and vehicle design, it is not clear to what extent current in-depth study procedures employ the data collection protocols and has access to the information required to produce in-depth studies that are helpful for ADAS design and evaluation. There were two aims for the present study. The first was to try to identify underlying contributing factors and causation patterns in a set of fatal crashes, based on reports from the road authorities’ in-depth investigation teams. The second aim was to assess the extent to which these factors are relevant for ADAS design and development, i.e. whether the reports produces contain the information necessary to perform ADAS relevant causation pattern analysis. To limit the study scope, the analysis was performed on intersection crashes only. Intersection crashes were selected because it is a crash type for which few, if any, ADAS have been deployed due to technical and other challenges involved, and a more precise problem description would be very helpful to their development (Bärgman and Smith, 2009). Furthermore, the study scope was limited to motor-vehicle crashes, i.e. crashes with vulnerable road user crashes (VRU’s, e.g. pedestrians and cyclists) were excluded, while crashes with Powered Two Wheelers (PTW), such as motorcycles and mopeds, were retained. VRU’s were excluded since they move and behave differently than motor vehicles; they use different parts of the road space, the kinematics are different, threat perception is made on different grounds, etc. A causation analysis which combines motor vehicle drivers with VRU’s therefore risks “hybrid” findings that may not be representative of either group. 2. Methodology The in-depth investigation reports used in the study come from Norway. In Norway all fatal crashes are analyzed in-depth by multidisciplinary crash investigation teams organised by the Norwegian Public Roads Administration (NPRA). The NPRA teams collect a mixture of data from on-the-scene investigations (i.e. right after the crash when the vehicles still are in place, mainly based on what the police observes) and/or on-the-site investigations (members of the team visit the crash site after the scene has been cleared to
Table 1 Vehicle sizes involved in the 28 fatal intersection crashes. Vehicle size Small Medium Large
Motorcycle Passenger car or similar Truck
Total
All
Turning
Going Straight
17 30 10
0 26 2
17 4 8
57
28
29
collect additional data). Also, records from later police interviews with the involved parties are collected. When all data is compiled, they produce an investigation report for each crash. The data for this study includes all 28 fatal motor vehicle intersection crashes that occurred in Norway in the years 2005–2007. In this sense, the sample is nationally representative of the crash type. However, 28 is a relatively low number given that there were 559 fatal crashes in total during that time period. Fatal crashes can thus be said to largely occur outside intersections, something which is further illustrated in Fig. 1 (data from Statistics Norway, 2011). This figure shows the relative frequency distributions over the major crash types for motor vehicle crashes and their associated fatalities and injuries. While crash and injury frequencies show similar distributions among crash types, fatalities have a rather different distribution. The implications of this in terms of result generalizability are further discussed Section 4. The data available for each crash varied in scope; for some crashes only the final report was available (usually a 5–10 page document), whereas for other crashes various protocols filled out during the investigation were also available. All crashes involved two vehicles except one, which involved three. There were three car-to-car crashes, whereasthe other 25 occurred between vehicles of pronounced size difference (i.e. PTW-to-car, car-to-truck). In Table 1, an overview of the involved vehicles is shown, along with information on to what extent vehicles in each size group was turning or going straight through the intersection. 2.1. Analysis procedure Prior to the analysis of contributing factors, the 28 cases were sorted following an intersection crash typology developed in a
784
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
LTAP/OD
LTAP/LD
SCP
Le Turn Across Path/ Opposite Direcon
Le Turn Across Path/ Lateral Direcon
Straight Crossing Paths
LTIP
RTIP
Le Turn Into Path
Right Turn Into Path
Fig. 2. Intersection crash types as defined by Najm et al. (2001).
Table 2 Distribution of crash types in analyzed material according to the typology from Najm et al. (2001). Fatal intersection accident conflict patterns LTAP-OD
LTAP-LD
SCP
RTIP
LTIP
Total
13
10
1
2
2
28
statistical study to describe crossing path crashes in the US (Najm et al., 2001). The typology sorts crashes based on actual and intended vehicle trajectories prior to the crash (Fig. 2). As an example, the typical scenario for a Left-Turn-Across-Path/Opposite Direction (LTAP/OD) crash is a left turning vehicle cutting across the path of another vehicle coming from the opposite direction, and which intends to cross the intersection on a straight path. This typology was used because it divided the US intersection crash population into subgroups that were clearly differentiated in terms of the typical circumstances under which they occur (Najm et al., 2001). Thus it might be expected that they could be differentiated in terms of contributing factors and causation patterns as well. Moreover, using vehicle trajectories as a typology basis provides a very natural frame of reference and principles for separation for the technical side of ADAS development, as vehicles on turning trajectories likely will have different sensor and Human Machine Interaction (HMI) requirements than vehicles going straight through an intersection. The distribution of the 28 crashes according to the typology is presented in Table 2. The most frequent conflict patterns are LTAPOD and LTAP-LD, which also were among the top three in Najm et al. (2001). However, the most common pattern found by Najm et al. (SCP crashes) is nearly absent from this sample. This is somewhat surprising, as there is no obvious reason why fatal crashes should be differently distributed in these conflict patterns compared to crashes in general. Unfortunately the Norwegian crash
typology can separate vehicles based on whether they came from the same or different roads but not based on whether they intended to go straight or turn. This means that LTAP-LD crashes cannot be separated from SCP crashes, which makes a general US–Norway comparison impossible. When the cases had been sorted, details on crash contributing factors for each driver in each in-depth case were coded using the Driving Reliability and Error Analysis Method (DREAM 3.0) (Wallén Warner et al., 2008). DREAM is an adaptation to the traffic safety domain of the Cognitive Reliability and Error Analysis Method (CREAM) (Hollnagel, 1998). The first version (Ljung, 2002) was developedin the FICA project (Ljung et al., 2007). It has since then been iteratively refined. The version used here was developed and applied in the European project SafetyNET (SAFETYNET, 2008), where approximately 1000 in-depth investigated crashes were DREAM coded. DREAM has two purposes: to support systematical classification of accident causation information, and to facilitate aggregation of that information into patterns of contributing factors (i.e. causation charts).1 DREAM includes three main elements: an accident model, a classification scheme and a method for applying the scheme to investigated events. The accident model serves as a framework and theoretical anchor for the classification scheme and method. It uses the human–technology–organisation (HTO) triad as a reference – represented by the driver (human), the vehicle and traffic environment (technology) and the organisation. The driving task is characterized through the Extended Control Model (ECOM; Hollnagel and Woods, 2005), which defines driving as a control task which includes working towards multiple parallel
1 It is worth pointing out that DREAM (like most other tools for accident analysis) is an organiser of explanations – not a provider. For any classification code in DREAM to be applicable, it must be supported by empirical information. If no information exists, there is nothing to classify.
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791 Table 3 Phenotypes in DREAM 3.0. Phenotypes
Specific phenotypes
Timing Speed Distance Direction Force Object
Too early action; too late action; no action Too high speed; too low speed Too short distance Wrong direction Surplus force; insufficient force Adjacent object
goals on different time scales. Accidents happen when an unsuccessful interplay between driver, vehicle and traffic environment leads to loss of control in one or more control processes, and that control cannot be regained given available time and/or resources. The classification scheme comprises a number of observable effects in the form of human actions and system events called phenotypes. Phenotypes are expressed in the general dimensions of time, space and energy, and consist of the following (Table 3). It also contains a number of possible contributing factors which may have brought about these observable effects. These are called genotypes, and are organised according to the driver–vehicle/traffic environment–organisation triad mentioned above. The driver category contains cognitive function genotypes related to problems with observation, interpretation and planning (in accordance with ECOM). It also includes more general person related factors that may degrade driving performance and hence contribute to an accident (e.g. fatigue). The other categories contain vehicle, traffic environment and organisation related contributing factors. The genotypes include contributing factors both at the sharp end (close in time/space to the crash) as well as at the blunt end (more distant in time/space, yet important for the development of events). In DREAM version 3.0 which was used here, genotypes are divided into 16 main categories, each belonging to one of four main groups: Driver, Vehicle and Traffic environment and Organisation (see Table 4). The phenotype/genotype distinction is adapted from biology, and refers to the fact that while for example all humans look more or less different (each phenotype is unique), we have share an identical genotype (a pool of genes), and differences between individuals depend on which combination of genes that was active in the shaping of that phenotype. Vehicle crashes can be viewed analogously, i.e. while each crash is unique, there is a common set of possible contributing factors that can be used to analyze (most) crashes. DREAM also includes a link system which specifies possible interactions between contributing factors. When case information on causation is coded into a chart, the link system ensures that the description of how one contributing factor leads to another is not arbitrary. The link system basically limits the range of possible factor interactions to those currently supported by scientific knowledge, thus restricting and guiding the coding of causation information. The inherent structure in the link system also makes it possible to aggregate causation information from multiple case studies in a structured, and principally semi-automated fashion, reducing the number of subjective judgements necessary to identify a pattern of contributing factors for a group of crashes. Naturally, the link system can be updated as new knowledge is gained. For a discussion of how to create and interpret aggregated causation charts using DREAM, see (Sandin and Ljung, 2007; Sandin, 2008). Lastly, the method for how to do the actual classification contains a detailed step-by-step instruction, as well as several stop rules. The latter are well defined conditions that determine when the analysis should come to an end. These are necessary, as the classification scheme represents a network rather than a hierarchy, and deciding when to stop the analysis can become arbitrary in the absence of such rules.
785
An assumption underlying the coding was that each driver has his/her own reasons for failing to adapt to the driving situation. Contributing factors were therefore coded separately for each involved driver, resulting in one causation chart per involved driver. Next, the individual charts for each crash type were aggregated based on which conflict trajectory (Fig. 2) the driver was on. Thus, two aggregate causation charts were created per crash type; one for turning drivers and one for drivers going straight. For the SCP crash, in which both drivers were going straight, only one aggregated chart was created. All coding was performed by the first author of this paper. 3. Results In the figures below, the total number of times a contributing factor occurs is represented by the number in brackets within each box. Note that the way DREAM is structured allows attribution of, for example, multiple missed observations to a single driver, depending on how many factors contributed to the missed observation. The occurrence frequency for some contributing factors might therefore exceed the number of aggregated causation charts. For visual guidance when looking for patterns, the factor frequency numbers are also indicated through box border thickness. For links between boxes, the number of times a link occurs is not written out, but indirectly represented through the thickness of the connecting arrows. Note also that to make the aggregated charts readable, the detailed information from the individual charts which motivates the choice of each phenotype and genotype has been removed. However, in the analysis, the detailed information from the individual charts was still used. The full analysis of each crash can be found in the final project report where the analysis was carried out (Akhtar et al., 2010). 3.1. LTAP/OD crashes The first analyzed crash type is Left Turn Across Path/Opposite direction (LTAP/OD) crashes. As described above, drivers going straight and turning drivers were put in separate groups prior to the aggregation. Below, the aggregate causation patterns for each of these groups are presented (Figs. 2 and 3), and the details of the most common contributing factors are described. 3.1.1. Left-Turn-Across-Path (LTAP) drivers While there were 13 LTAP-OD crashes in the sample, there are 14 phenotypes (and hence 14 LTAP drivers) described in Fig. 3. This is because one of the crashes involved three vehicles, and two of these were turning. A majority of the LTAP drivers involved in LTAP-OD crashes were coded with the phenotype Timing: too early (13 of 14). This indicates that they began to turn through the intersection before it was appropriate to do so. Here, “before it was appropriate” refers to the fact that these turning drivers were expected to yield to any vehicle on a straight crossing path before turning, which they in these cases did not do. Two of those too early turns are preceded by Misjudgment of time gaps, i.e. the turning driver overestimates the time available for completing the turn before the oncoming vehicle reaches the intersection. The other 11 are coded with Misjudgment of situation as the main contributing factor. Contributing to those misjudgments are 8 instances of Missed observation, meaning that the driver did not see the other vehicle at the time when s/he decided to initiate the turn. A closer reading of the individual charts shows that the conflict vehicle in 7 of those 8 instances was a PTW, which the drivers were unable to perceive at “checkpoint time”, i.e. when they decided it was okay to make the turn. Reasons for not perceiving the vehicles
786
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
Table 4 Genotypes in DREAM 3.0. Genotypes (B–Q) Human (B–F)
B: Observation Missed observation (B1) Late observation (B2) False observation (B3)
Technology (G–M) Vehicle (G–I)
Traffic environment (J–M)
G: Temporary HMI problems Temporary illumination problems (G1) Temporary sound problems (G2) Temporary access limitations (G4)
J: Weather conditions Reduced visibility (J1)
N: Organisation Time pressure (N1)
Strong side winds (J2)
Incorrect ITS-information (G5)
K: Obstruction of view due to object Temporary obstruction of view (K1) Permanent obstruction of view (K2)
Irregular working hours (N2) Heavy physical activity before drive (N3) Inadequate training (N4)
C: Interpretation Misjudgement of time gaps (C1)
H: Permanent HMI problems
Misjudgement of situation (C2)
Permanent illumination problems (H1) Permanent sound problems (H2) Permanent sight obstruction (H3)
D: Planning Priority error (D1)
E: Temporary personal factors
Organisation (N–Q)
I: Vehicle equipment failure
L: State of road Insufficient guidance (L1) Reduced friction (L2) Road surface degradation (L3)
Equipment failure (I1)
Object on road (L4)
Fear (E1)
Inadequate road geometry (L5)
Inattention (E2) Fatigue (E3) Under the influence of substances (E4) Excitement seeking (E5) Sudden functional impairment (E6) Psychological stress (E7)
M: Communication Inadequate transmission from other road users (M1) Inadequate transmission from road environment (M2)
O: Maintenance Inadequate vehicle maintenance (O1) Inadequate road maintenance (O2) P: Vehicle design Inadequate design of driver environment (P1) Inadequate design of communication devices (P2) Inadequate construction of vehicle parts and/or structures (P3) Unpredictable system characteristics (P4) Q: Road design Inadequate information design (Q1) Inadequate road design (Q2)
F: Permanent personal factors Permanent functional impairment (F1) Expectance of certain behaviours (F2) Expectance of stable road environment (F3) Habitually stretching rules and recommendations (F4) Overestimation of skills (F5) Insufficient skills/knowledge (F6)
going straight include Reduced visibility (here: MC driver wearing dark clothing against dark background), Inadequate transmission from other road users (here: sub-standard MC lighting) and various obstructions to view. The latter include obstructions inside the vehicle (Temporary sight obstruction, here items on the dashboard in one case and a dirty windshield in the other), and obstructions outside the vehicle. For outside distractions, Temporary obstruction of view indicates blocked lines of sight due to other vehicles in the traffic environment at critical decision moments, while the objects coded with Permanent obstruction of view primarily refers to vegetation. There is however also one instance of a traffic sign blocking the line of sight, and one instance where the road lighting design actually reduced object visibility in rain. Apart from difficulties in perceiving the road user going straight, the turning drivers also have 8 instances of Inattention coded as contributors to Misjudgment of situation. These instances of inattention are largely attributed to Driving-related distractors outside vehicle (6 instances), which means that the drivers were focused on negotiating some other relevant task in the traffic environment. In four of the six instances, this involved tracking a third vehicle also about to enter the intersection. The fifth instance involved tracking pedestrians who are crossing the road somewhat randomly (outside a school), and the last involved maneuvering a large truck and trailer combination through a (for that vehicle combination) tight turn.
Another frequent contributor to Missed observation is Expectance of certain behaviours, with four instances. Three of these refers to drivers who expected other vehicles to keep approximately to posted speed limits and who therefore did not look sufficiently far down the road to discover a MC travelling at very high speed.
3.1.2. OD drivers There were 13 drivers on a straight crossing path in the LTAP-OD crashes, and the aggregated causation chart for these is shown in Fig. 3. The phenotype coding for these drivers is more diverse than for the turning drivers. 5 are coded as Timing: no action, indicating that they did not perform any type of steering and/or braking prior to the crash. The 4 coded as Timing: too late have to some extent started a corrective action, but to late to avoid the crash. 3 drivers are coded as Speed: too high. Two involve MCs travelling at substantially higher speeds than legally permitted. The third case involves a car which negotiates a curve with very limited visibility at slightly above legal speed, which leaves the driver with insufficient time to respond to a vehicle standing still at an intersection just after the curve. The last driver is coded as Force: insufficient force, which here means that if the driver had used the normal braking capacity of the MC he was driving the crash most likely would have been avoided (Fig. 4).
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
787
Fig. 3. Aggregated causation chart for the 14 drivers involved in LTAP/OD crashes who were on a LTAP trajectory.
The most common contributing factor for OD drivers is Misjudgment of situation (12 instances), 10 of which are preceded by Expectance of certain behaviours. Here this indicates that the drivers going straight were expecting any turning vehicles to yield, i.e. the drivers have most likely seen the other vehicle, but expects it to stop/wait for them. There are also three drivers for whom the chain Inadequate training → Insufficient skills/knowledge leads to Misjudgment of situation. Two of these lacked a drivers license for the MC they were driving. These two were also coded as Under the influence of substances, as well as with the chain Excitement seeking → Priority error. The third driver was on a high performance MC, the full brake capacity of which he did not master. 3.2. LTAP/LD crashes The second analyzed crash type is Left Turn Across Path/Lateral Direction (LTAP/LD) crashes. As described above, drivers going straight and turning drivers were put in separate groups prior to the aggregation. Below, the aggregate causation patterns for each of these groups are presented (Figs. 6 and 7), and the details of the most common contributing factors are described. 3.2.1. Left-Turn-Across-Path (LTAP) drivers The aggregated causation pattern for the 10 turning drivers in LTAP-LD crashes is shown below in Fig. 5. As can be seen in the chart, the turning drivers in LTAP-LD crashes are, just like the turning drivers in LTAP-OD crashes, almost exclusively phenotype coded as Timing: too early (9 of 10). As for LTAP-OD crashes, this indicates that they began to turn through the intersection before it was appropriate to do so. The immediate antecedent for all 10 crashes is Misjudgment of situation. Contributing to that misjudgment are 7 instances of Missed observation, to which in turn a combination of Permanent obstruction to view and Inattention contributes. As above, the instances of Permanent obstruction to view mainly refer to
vegetation elements in the intersection (6 of 7 instances), but traffic environment elements also play a part (2 signposts and one case of inadequate intersection layout). The 6 instances of Inattention are mainly attributed to distractors, the most common being the 4 instances of Driving related distractors outside vehicle, e.g. other vehicles which must be considered when negotiation of the traffic situation. There is also one instance of a Driving related distractor inside vehicle and one instance where Inattention is attributed to the driver being drunk (Alcohol). In two of the three cases where Insufficient skills/knowledge has been coded, this refers to young (18 years) and relatively inexperienced drivers who have had their driving licenses for a short time. The third case involves a driver who is negotiating the manual shift of a rental car, while being used to automatic gear in his own vehicle. 3.2.2. OD drivers The aggregated causation pattern for the 10 drivers going straight in LTAP-LD crashes is shown below in Fig. 6. Five of these are coded as Timing: no action, which means that they did not perform any type of steering and/or braking prior to the crash. The 3 coded as Timing: too late have initiated corrective actions, but too late to avoid the crash. One driver is coded as Speed: too high, indicating a travel speed substantially over the speed limit. In terms of contributing factors, the most common contributor is Misjudgment of situation. As for OD drivers in LTAP-OD crashes, a majority of these are in turn attributed to Expectance of certain behaviours, i.e. the drivers expects non-priority vehicles to yield. There are also 2 instances of Missed observation and 2 instances of Late observation contributing to Misjudgment of situation. The instances of Missed observation are in turnattributed to one instance of Inadequate transmission from other road users and one instance of Permanent obstruction to view. There is one instance of Misjudgment of time gaps, where an MC driver overestimated the time available for passing through the intersection before the turning vehicle would be blocking the MC’s
788
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
Fig. 4. Aggregated causation chart for the 13 drivers involved in LTAP/OD crashes who were on an OD trajectory.
travel path. This driver is also coded as Excitement seeking → Priority error and Alcohol → Under the influence of substances and Speed: too high.
as trying to beat a third vehicle (coming from his right) to the intersection by driving fast (the conflict vehicle came from his left). 3.4. Right-Turn-Into-Path (RTIP) crashes
3.3. SCP crashes There was just one SCP crash in the material. Here, both drivers were coded with Timing: no action, indicating that neither made any attempt to stop or yield as they approached the intersection. Both drivers had difficulties seeing the other vehicle (Late observation and Missed observation respectively), due to a mutual Permanent obstruction of view. One driver is also coded as Excitement seeking → Priority error in combination with Driving-related distractors outside vehicle. Here, this refers to the fact that the driver was reported
There were two RTIP crashes in the sample. Of the drivers turning right, one is coded with Timing: too early, indicating that he commenced the turn before the intersection was clear, while the other is coded as Timing: no action, indicating that he drove into the intersection without yielding to the vehicle going straight. Contributing factors are one instance of Misjudgement of situation due to Missed observation and Expectance of certain behaviourism, which refers to one of the drivers who was not looking for vehicles travelling at speeds much higher than posted speed limit, and who
Fig. 5. Aggregated causation chart for the 10 drivers involved in LTAP/LD crashes who were on a LTAP trajectory.
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
789
Fig. 6. Aggregated causation chart for the 10 drivers involved in LTAP/LD crashes who were on a LD trajectory.
therefore did not look sufficiently far down the road to see the MC approaching at high speed. Of the two drivers going straight one was speeding on a motorcycle but expected the turning driver to yield. In the other case, the right turning car just pulled out in front of a semitrailer, and the trailer driver did not have time to react at all. 3.5. Left-Turn-Into-Path (LTIP) crashes There were two LTIP crashes in the sample. Both turning drivers commenced the turn prematurely, (Timing: too early) due to Missed observation (s/he did not observe an oncoming MC). In both cases, the approaching MC was travelling much faster than the speed limit and was difficult to perceive in terms of optical visibility. The MC drivers on a straight path in these crashes both were driving much faster than legally permitted (both coded as Excitement seeking→Priority error). One of the drivers was involved in what can be referred to as a peer pressure situation, i.e. wanting to show off to a friend (Psychological stress), and overestimating his own driving skills (Overestimation of skills). 4. Discussion There were two aims in the present study. The first was to try to identify underlying contributing factors and causation patterns relevant for ADAS design and development in intersection crashes, based on reports from a set of in-depth investigated fatal crashes. The second aim was to evaluate whether the data collection and documentation procedures used to perform the in-depth investigations yield the information necessary to perform this type of causation pattern analysis. In terms of the first objective, it is clear that there is a lot of relevant information available in these in-depth fatal crash reports. Also, separating drivers based on trajectory following the intersection conflict typology developed by Najm et al. (2001) is useful in many aspects. In the analysis performed here however, it seems warranted to aggregate also at a higher level, based on whether drivers are going straight or turning. In other words, when looking at the causation charts, it seems like turning drivers have contributing factors in common across the Najm conflict types, as do drivers going straight, and there seems to be a clear difference between these two groups. The patterns in each respective group, and what the implications of those are for ADAS design, will be discussed below. After that, the discussion will come back to the data quality issue, i.e. the relationship between data collection and documentation procedures and analysis outcome.
4.1. Common factors for turning drivers Turning drivers are almost uniformly coded with timing: too early, indicating that they initiate the turn before the crossing vehicle has passed. In terms of ADAS development, this clearly indicates a need for turning decision support, i.e. smart ways of helping the driver understand when it is safe to initiate a turn and when it is not. Moreover, the misjudgements of situation underlying these too early turns seems to be largely driven by two types of problems. One is that the drivers do not see the conflict vehicle at the time of making the decision to turn, due to physical obstructions to view. In terms of ADAS design, this means that regular, vehicle based, sensors such as cameras and radar will not be sufficient to solve the problem, as their lines of sight would be hindered as well. Instead, vehicle-to-vehicle communication technology probably needs to be considered. Second, a large number of the turning drivers are focused on some other part of the traffic situation than the oncoming vehicle (Driving-related distractors outside the vehicle). This means than an information/warning ADAS not only needs to inform the driver about a problem, it needs to do so in a way that “breaks through” the drivers’ current focus, i.e. the ADAS has to redirect attention rather than just get attention. Consequently, evaluation of an information/warning ADAS for turning drivers at intersections needs to be performed in a situation where the test drivers really are engaged in some concurrent, high priority, driving related task (like tracking another vehicle about to enter the intersection). If this condition is not met, the ADAS ability to redirect the drivers’ attention will not be properly assessed. This is further underlined by the fact that turning drivers are also frequently coded with Expectance of certain behaviours, referring to turning drivers not adjusting their regular intersection scanning pattern to accommodate vehicles travelling at speeds well above the speed limit. In other words, the behaviour of the conflict vehicle lie outside what normally can be expected in many of the cases analyzed here. This circumstance also needs replicating in an evaluation situation. 4.2. Common factors for drivers going straight Drivers going straight show a high frequency of planning failures due to Expectance of certain behaviours. This likely reflects the fact that drivers going straight often have the right of way and therefore expect turning vehicles to yield. Adding to this interpretation is the fact that drivers going straight are much less frequently coded with Missed observation and Obstructions to view than turning drivers. Given that the driver going straight has the right of way, the task of
790
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
identifying and responding to a conflict vehicle in an intersection rests with the turning driver. Sight limitations are thus more debilitating for the turning driver’s performance, and hence more likely to be reported as a contributing factor for them than for drivers going straight. In terms of ADAS development and evaluation, this means than an information/warning ADAS not only needs to inform the driver going straight about a problem, it needs to do so in a way that convinces a driver who believes s/he has the right of way to actually brake and/or swerve. This sets a very challenging target for the ADAS threat assessment. First, for a driver going straight to trust and rely on such an ADAS it has to be right most of the time, i.e. being perceived by the driver as making relevant and warranted calls rather than crying wolf. Second, the threat assessment needs to be carried out sufficiently well ahead in time to allow the driver to act on a warning. It is worth noting that the group of drivers going straight contains all 17 PTW drivers involved in the 28 fatal crashes, many of which were driving well above the speed limit. Also, when it comes to contributing factors like Excitement seeking and Alcohol/Drugs, as well as Overestimation of skills and Insufficient skills/knowledge, these are almost exclusively coded for drivers going straight. In terms of ADAS evaluation and development, this presents some interesting strategy choices. If one wants to develop an ADAS that would work for all drivers going straight, this means that the test drivers at evaluation somehow need to represent the conditions which the coded factors indicate (excitement seeking, alcohol, etc.) for the assessment to be correct. This presents interesting challenges for test subject preparation and evaluation. It is also challenging for the ADAS technical capacity, considering the actual travel speeds that needs to be handled. If one on the other decides to develop an ADAS for the “normal” driver, i.e. not speeding, untrained, etc., then the number of drivers which the ADAS actually can help may be quite limited. 4.3. Data quality A general observation is that there are in general fewer contributing factors coded for drivers going straight than for turning drivers. In other words, despite the investigators being trained observers tasked with taking an objective view of the data, they seem more likely to provide deeper and fuller explanations of why the turning driver gets into trouble, compared to the driver going straight. While it might be that there simply is less relevant information to code for drivers going straight, a more likely reason for this focus on the turning drivers (albeit involuntary) is connected to the fact that the turning driver usually is the one held legally liable for the crash. Since the driver going straight normally has the right of way it is easy to conclude that it was not his/her fault, and investigation might therefore come to focus on explaining why the driver at legal fault ended up in the crash. An example which illustrates this asymmetry is the number of reported obstructions to view due to signposts and vegetation. Since these are part of the traffic environment, i.e. the infrastructure, one would assume that any blockage in lines of sight is reciprocal, i.e. if driver A cannot see driver B, the reverse should also be true. However, while external obstructions to view are frequently reported as contributing for turning drivers, they are rarely reported as contributing for their counterparts in the same crash. This investigator bias towards focusing on the party held liable presents a challenge for ADAS development. It is true from a physics standpoint that if two vehicles are on a colliding path, it will almost always be easier for one of the vehicles to perform the avoidance maneuver necessary to prevent a collision. However, it is far from clear that this vehicle necessarily belongs to the driver at fault; it is
a kinematic relationship between masses rather than a moral relationship between operators. In other words, who to blame is usually irrelevant from an ADAS development point of view. Underreporting of contributing factors for one of the parties involved based on reasoning about guilt thus hinders rather than helps countermeasure development. This underlying investigator mind-set, if it can be called that, would need addressing to avoid future bias in reported information. The amount of available data in this study varied somewhat between crashes; for some there was just the final report, whereas for other accidents various protocols filled out by the investigators during the investigation were available. The latter extra information several times proved to be very valuable for the analysis, not in the sense that it brought new information to light, but in order to rule possible contributing factors out. While the final reports overall are very good at compiling the relevant information from the other documents, they describe inclusions rather than exclusions. In other words, the reasons for why certain factors are thought to contribute are included, but the reasons for excluding other possible factors are left out. Here, the extra information in the other protocols could at times be used to discard certain possible contributing factors, the established absence of which certainly made a difference in terms of how the accident causation process was reconstructed. An interesting fact is that contributing factors often are pointed to as main contributors to fatal crashes, e.g. high speed, drugs/alcohol and inadequate driver training, played a role in 12 of 28 accidents. Two implications follow from this. First, if one were to evaluate an ADAS addressing intersection crashes, finding and preparing a relevant group of test persons would be a challenge, at least if the group is meant to be representative of the drivers involved in the crashes investigated here. For example, using intoxicated test persons has numerous ethical issues associated with it. Second, while this confirms the prevalence of these factors, it also indicates that most drivers end up in these situations due to combinations of less auspicious contributing factors. In this study, a number of the possible contributing factors that the DREAM methodology contains were never applied to any of the 28 crashes. As DREAM has been put through validation work and corroboration with other researchers’ findings on possible crash causes, there is reason to further investigate why many of the contributing factors available in DREAM never were applied. There are at least three possible explanations for this. One is that the information necessary for coding those genotypes usually comes from detailed driver interviews. As this is inherently impossible for at least one subject in a fatal crash, there is a risk that the analysis becomes systematically biased. However, unless contributing factors for the fatally wounded systematically differ from the factors for those who survived, the survivors should still provide information related to these factors if present. Another possible explanation is that the Norwegianin-depth investigation teams rarely conduct their own driver interviews. Rather they rely on information from the police interviews. Here it can be hypothesised that crash survivors are not always completely forthcoming when describing crash circumstances to the police. A third possibility is that the full set of contributing factors in DREAM might not be necessary for analysis of fatal crashes, i.e. the factors not coded are simply not relevant contributors in crashes with fatal outcomes. Which of these explanations are the best is for future studies. As all coding was performed by the author, the question of whether another coder would have come to different conclusions naturally arises. While assessing this topic was outside the scope of this study, a previous interrater reliability study showed a relatively high interrater agreement for coders applying DREAM to a set of sample cases (Wallén Warner and Sandin, 2010). Thus, while some variation must be expected if another coder would code the
M. Ljung Aust et al. / Accident Analysis and Prevention 45 (2012) 782–791
same set of cases, the variability could be expected to stay within reasonable limits. This is however also a topic for future study. A final topic for discussion is the representativity of the contributing factors found for the 28 crashes studied here. On one hand, since the data set contains all crashes of the selected type that occurred in the selected years, the data set is clearly nationally representative. On the other hand, as Fig. 1 shows, fatalities have a different distribution across crash types than crashes and injuries in general. To extend the conclusions on contributing factors drawn for the data set here to crashes in general, one thus would need to explain the distribution difference in a way that does not involve the causation analysis. One possibility is to argue that the distribution difference reflects the energies involved rather than what brought the crash about. If so, Fig. 1 could be interpreted as showing that impact speeds are highest in head-on collisions and lowest in rear-end crashes, with single vehicle crashes and intersection crashes in between. However, while this explanation seems plausible, it does not a priori rule out the possibility that fatal crashes have at least somewhat different contributing factors than crashes in general. For example, if the numbers on high speed, drugs/alcohol and inadequate driver training above were representative of all crashes, then they would play a role in 4 of 10 intersection crashes, which seems quite high. Finding a way of deciding between these explanations is clearly a topic for future study. 5. Conclusions There were two aims in the present study. The first was to try to identify underlying contributing factors and causation patterns relevant for ADAS design and development in intersection crashes, based on reports from a set of in-depth investigated fatal crashes. The second aim was to evaluate whether the data collection and documentation procedures used to perform the in-depth investigations yield the information necessary to perform this type of causation pattern analysis. The study shows that it is indeed possible to identify such causation patterns in this type of data. This study indicates that turning drivers to a large extent are faced with perception difficulties and unexpected behaviours in relation to the conflict vehicle, while at the same time trying to negotiate a demanding traffic situation. Drivers going straight on the other hand have less perception difficulties. Instead, their main problem is that they largely expect turning drivers to yield. When this assumption is violated, they are either slow to react or do not react at all. These findings are very informative when it comes to development and evaluation of ADAS intended to address intersection crashes, as they present direction and place constraints on how the ADAS needs to interact with the driver to be successful. Furthermore, contributing factors often pointed to in literature, e.g. high speed, drugs and/or alcohol and inadequate driver training, played a role in 12 of 28 accidents. While this confirms their prevalence, it also indicates that most drivers end up in these situations due to combinations of less auspicious contributing factors. In terms of data collection and documentation, the analysis shows that there seems to be an (involuntary) analyst focus on
791
identifying contributing factors for the driver legally held liable, which in this crash type often is the turning driver, while less attention is paid to the driver not at fault (the driver going straight). However, since who to blame often is irrelevant from a countermeasure development point of view, this underlying mind-set needs addressing to avoid future bias in the reported information. Acknowledgements This study draws on accident investigations performed by the Norwegian Public Roads Administration, and was funded by the Research Council of Norway, under the RISIT programme (Risk and Safety in Transport). References Akhtar, J., Ljung Aust, M., Eriksson, R.J., Fagerlind, H., Høye, A., Phillips, R., Sagberg, F., 2010. Factors Contributing to Road Fatalities: Analysis of In-depth Investigation Data from Passenger Car Intersection Crashes and From Collisions Between Bicycle and Motorized Vehicle. Oslo: Institute of Transport Economics, Oslo, Norway, TØI Report 1067. Bärgman, J., Smith, K., 2009. Ivss Intersections Accidents: Analysis and Prevention. Final Report. IVSS Programme, Borlänge, Sweden. Hollnagel, E., Woods, D.D., 2005. Joint cognitive systems: foundations of cognitive systems engineering. CRC Press, Boca Raton, FL. Hollnagel, E., 1998. Cognitive Reliability and Error Analysis Method (CREAM). Elsevier Science. Larsen, L., 2004. Methods of multidisciplinary in-depth analyses of road traffic accidents. Journal of Hazardous Materials 111, 115–122. Ljung, M., 2002. DREAM: Driving Reliability and Error Analysis Method. Master Thesis. Linköping University. Ljung, M., Fagerlind, H., Lövsund, P., Sandin, J., 2007. Accident investigations for active safety at chalmers – new demands require new methodologies. Vehicle System Dynamics 45 (10), 881–894. Midtland, K., Muskaug, R., Sagberg, F., Jørgensen, N.O.C., 1995. Evaluation of the In-depth Accident Investigations of the Swedish National Road Administration. Najm, W.G., Mironer, M., Koziol, J., Wang, J.S., Knipling, R.R., 1995. Synthesis Report: Examination of Target Vehicular Crashes and Potential Its Countermeasures. U.S. Department of Transportation, Volpe National Transportation Systems Center, Cambridge. Najm, W.G., Smith, J.D., 2002. Year Breakdown of light vehicle crashes by pre-crash scenarios as a basis for countermeasure development. In: Proceedings of the Proceedings of the International Conference on Applications of Advanced Technologies in Transportation Engineering, pp. 712–719. Najm, W.G., Smith, J.D., Smith, D.L., 2001. Analysis of Crossing Path Crashes U.S. Department of Transportation. John A.Volpe National Transportation Systems Center, Cambridge. OECD,1988. Road accidents: on-site investigations. In: Development. O.F.E.C.-O.A., Paris. SAFETYNET, 2008. Deliverable 5.8: In-depth Accident Causation Database and Analysis Report. European Road Safety Observatory. Sandin, J., 2008. Aggregating Case Studies of Vehicle Crashes by Means of Causation Charts – An Evaluation and Revision of the Driving Reliability and Error Analysis Method. Chalmers University of Technology. Sandin, J., Ljung, M., 2007. Understanding the causation of single-vehicle crashes: a methodology for in-depth on-scene multidisciplinary case studies. International Journal of Vehicle Safety 2 (3), 316–333. Statistics Norway, 2011. Tabell 03443: Personer drept eller skadd i veitrafikkulykker, etter trafikantgruppe og ulykkesgruppe (avslutta serie). Statistikkbanken område 10.12: Transport og kommunikasjon. Statistisk Sentralbyrå, Oslo. Wallén Warner, H., Ljung Aust, M., Björklund, G., Johansson, E., Sandin, J., 2008. Manual for DREAM 3.0, Driving Reliability and Error Analysis Method. Deliverable D5.6 of the EU FP6 Project SafetyNet. European Commission, Brussels. Wallén Warner, H., Sandin, J., 2010. The intercoder agreement when using the driving reliability and error analysis method in road traffic accident investigations. Safety Science 48 (5), 527–536.