Requirements of a system to reduce car-to-vulnerable road user crashes in urban intersections

Accident Analysis and Prevention 43 (2011) 1570–1580

Contents lists available at ScienceDirect

Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap

Requirements of a system to reduce car-to-vulnerable road user crashes in urban intersections Azra Habibovic ∗ , Johan Davidsson Division of Vehicle Safety, Department of Applied Mechanics, Chalmers University of Technology, SE-412 96, Gothenburg, Sweden

a r t i c l e

i n f o

Article history: Received 16 July 2010 Received in revised form 20 March 2011 Accepted 21 March 2011 Keywords: Intersection Vulnerable road user Crash data Causation factor Advanced driver assistance system Functional requirement

a b s t r a c t Intersection crashes between cars and vulnerable road users (VRUs), such as pedestrians and bicyclists, often result in injuries and fatalities. Advanced driver assistance systems (ADASs) can prevent, or mitigate, these crashes. To derive functional requirements for such systems, an understanding of the underlying contributing factors and the context in which the crashes occur is essential. The aim of this study is to use microscopic and macroscopic crash data to explore the potential of information and warning providing ADASs, and then to derive functional sensor, collision detection, and human–machine interface (HMI) requirements. The microscopic data were obtained from the European project SafetyNet. Causation charts describing contributing factors for 60 car-to-VRU crashes had been compiled and were then also aggregated using the SafetyNet Accident Causation System (SNACS). The macroscopic data were obtained from the Swedish national crash database, STRADA. A total of 9702 crashes were analyzed. The results show that the most frequent contributing factor to the crashes was the drivers’ failure to observe VRUs due to reduced visibility, reduced awareness, and/or insufficient comprehension. An ADAS should therefore help drivers to observe the VRUs in time and to enhance their ability to interpret the development of events in the near future. The system should include a combination of imminent and cautionary collision warnings, with additional support in the form of information about intersection geometry and traffic regulations. The warnings should be deployed via an in-vehicle HMI and according to the likelihood of crash risk. The system should be able to operate under a variety of weather and light conditions. It should have the capacity to support drivers when their view is obstructed by physical objects. To address problems that vehicle-based sensors may face in this regard, the use of cooperative systems is recommended. © 2011 Elsevier Ltd. All rights reserved.

1. Introduction Crashes involving vulnerable road users (VRUs) such as pedestrians and bicyclists constitute a large societal and economic cost. In the European Union, there are more than 300,000 VRU injuries and 4800 VRU fatalities every year, which correspond to approximately 16% of the total number of injuries and 19% of the total number of fatalities (ERSO, 2008). In the United States, VRUs accounted for approximately 113,000 or 4.5% of all road traffic injuries and 5352 or 13% of all road traffic fatalities in 2007 (NHTSA, 2008). On a world scale: a report by Singh (2005) shows that VRUs constitute a much larger fraction of all vehicle related fatalities and injuries in developing countries than in industrialized countries. A large fraction of VRU injuries and fatalities occur at urban intersections (about one third) and in collisions with passenger cars (about two thirds), see NHTSA (2008) and SIKA (2008).

∗ Corresponding author. Tel.: +46 31 772 8430; fax: +46 31 764 7188. E-mail address: [email protected] (A. Habibovic). 0001-4575/$ – see front matter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.aap.2011.03.019

These figures indicate that the safety of VRUs must be improved. In developing countries, this can be achieved by adopting the same basic safety principles that have resulted in a great reduction of injuries and fatalities in industrialized countries, such as adequate road design and traffic management. In industrialized countries, on the other hand, complementary countermeasures are needed to further reduce VRU injuries and fatalities. Active safety systems belong to such countermeasures. These systems can assist the VRUs or drivers, or both, in identifying threats and avoiding crashes. The present study focuses on the systems assisting drivers. These systems, known as advanced driver assistance systems (ADASs), can be roughly divided into four categories: information providing, warning/feedback providing, intervention with the driver in the loop, and intervention without the driver in the loop (Carsten and Nilsson, 2001). This study focuses on the first two categories. As explained later, this decision originates from limitations in the data sources used. The data sources used here contain sparse information on the emergency phase where the intervening systems are usually deployed.

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

To determine what functional requirements the information and warning providing systems should fulfill to address drivers’ real difficulties in car-to-VRU crashes, an in-depth understanding of why and how these crashes occur is necessary. Since most crashes are caused by a combination of behavioral, technological, and environmental factors (Ljung, 2002) it is also necessary to understand how contributing factors interact with each other. To acquire this understanding, two types of data are generally available. One is macroscopic crash data, usually consisting of police-reported crashes in databases such as STRADA (Sjöö and Ungerbäck, 2007). The other is microscopic data, normally in the form of in-depth crash investigations. Macroscopic data can often provide statistically significant information on context variables such as place and the number of road users involved, but contain little information about the sequence of events leading up to the crash. Microscopic data on the other hand, while rarely statistically significant, can often provide rich information on contributing factors, and be used to generate what may be called causation charts, depicting the contributing factors and their interactions for each crash (Ljung et al., 2007; Sandin, 2008). However, a causation chart from a single crash case does not allow the identification of recurrent causes; it cannot be used to develop active safety systems to counter a recurrent effect (Sandin, 2008). To be truly useful for defining ADASs, causation charts must therefore be aggregated. Ideally, macroscopic and microscopic data should be used in combination. However, the literature surveyed (e.g. Gavrila et al., 2003; Andreone et al., 2006) reveals that ADAS specifications are often based solely on macroscopic crash data, i.e. the information on the problem that the ADAS should resolve is obtained only from macroscopic crash data sources. Since this means that they most likely lack information on crash contributing factors, there is a risk that ADASs specified that way will not accurately address drivers’ support needs. This study aims to inform the design of ADASs that can prevent car-to-VRU crashes in urban intersections. First, the study explored a microscopic data source to acquire a deeper understanding of drivers’ support needs in relation to this crash type. Next, it examined the potential of information and warning providing ADASs to address these needs. Then, the study analyzed a macroscopic data source to form an understanding of the context for the crash type. Based on results of the macroscopic and microscopic data analyses, the functional requirements were derived for: (a) sensors, (b) a collision risk assessment algorithm, and (c) the human machine interface. 2. Method 2.1. Microscopic data The microscopic, or in-depth, crash data were collected by the European Commission project SafetyNet, see Björkman et al. (2008). The crash investigations were carried out by multidisciplinary teams, including traffic and vehicle engineers, as well as driver behavior experts. Data from 2005 to 2008 were collected for Italy, Finland, Germany, the Netherlands, Sweden, and the United Kingdom. The data collected include human, technology, and organization related parameters from crashes with at least one injury or fatality. The final analysis of each SafetyNet crash included the compilation of what can be referred to as a multi-linear causation chart. One chart per road user was produced, based on the assumption that each road user involved has individual reasons for failing to adapt to the development of events prior to a crash. In this context, “failing to adapt” or “adaptation failure” refers to the inability to successfully respond to changes in the driving conditions (e.g. failing to stop at the stop line when the traffic signal changes from green to red). Each

1571

chart describes how the identified contributing factors combine to produce the adaptation failure (and hence the crash) for that particular road user. The causation charts in the SafetyNet database were compiled according to the analytical method SafetyNet Accident Causation System (SNACS), Ljung (2007). The SNACS is a version of the Driver Reliability and Error Analysis Method (DREAM), see Ljung (2002), Ljung et al. (2007), Sandin (2009), Ljung Aust (2010), and Wallén Warner and Sandin (2010). The DREAM was developed by a Swedish project with aims similar to those of the SafetyNet project, i.e. to facilitate identification of crash contributing factors that may be prevented by using interactive (information and warning providing) ADASs. The DREAM was an adaptation, to the road safety domain, of the more generic Cognitive Reliability and Error Analysis Method (CREAM), Hollnagel, 1998. While the field of “human factors” has a long tradition of developing methods for human error modeling and analysis, few (if any) but DREAM have so far been explicitly adapted to the domain of road safety (for a review see Salmon et al., 2010). This is the main reason why the DREAM derivative SNACS was used in the SafetyNet project and in our study. Furthermore, while traditional crash analysis considers an adaptation failure to be an individual driver’s failure (Salmon et al., 2010), the SNACS treats an adaptation failure as a human-technologyorganization system failure. This approach to an adaptation failure is more suitable for the development of interactive crash avoidance systems. Compared with other crash analysis methods, the SNACS falls into the category of system based methods (Salmon et al., 2010), and has an underlying crash model that puts it in the family of complex linear system models (the current mainstream approach in human factors crash analysis methods, see Lundberg et al., 2009). The present study retrieved causation charts for all car drivers involved in car-to-VRU crashes registered in the SafetyNet database. In total, the database contains 128 car-to-VRU crashes, of which 60 were selected for further analysis. The crashes selected occurred in X intersections (35) or T intersections (25), and contain the causation charts for which the coders’ level of confidence was either high or reasonable. Of these 60 crashes, 39 are car-tobicyclist and 21 are car-to-pedestrian. In a SNACS chart, it is possible to distinguish between two elements: the critical event and contributing factors. The coding of the critical event captures the observable consequence of an adaptation failure to the traffic situation, such as failing to brake or to give way. These consequences are expressed in the physical dimensions of time, space, and energy. The contributing factors are those that can be identified in the crash data as contributing to the occurrence of the critical event. In the SNACS manual, the set of possible contributing factors available to the coder includes both sharp end factors (close in time/space to the crash) and blunt end factors (more distant in time/space from the crash, see Reason (1997) for details of this distinction). This set contains factors which are directly driver and vehicle related (e.g. fatigue or a tire blowout). There are also factors which refer to various procedures within the organizations and/or institutions that shape the conditions under which driving takes place. These include road and vehicle design, road and vehicle maintenance, and driver training. There are eight general critical event categories and 16 general contributing factor categories (Table 1). Each of them has a set of subcategories that can be used to code more detailed information. The contributing factors are linked to each other and to the critical event caused. As exemplified in Fig. 1, several causes can be linked to a critical event. However, there is only one critical event per chart. In terms of their definitions, some of the critical events are very closely related, e.g. distance, speed and timing are naturally interconnected. However, according to the SNACS coding manual (Ljung,

1572

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

Table 1 The main categories of critical events and contributing factors. Contributing factors

Critical events

Organization and traffic environment

Vehicle

Driver

J: Communication driver ↔ environment K: Maintenance – condition of road N: Design of traffic environment M: Organization

G: Temporary HMI problems H: Permanent HMI problems I: Equipment failure K: Maintenance – condition of vehicle O: Vehicle design

B: Observation C: Interpretation D: Planning E: Temporary personal factors F: Permanent personal factors J: Communication driver ↔ driver L: Experience or knowledge

2007), while several critical events may logically be possible when coding crash information, one of them is usually more appropriate given the data available. Basically, selecting the most appropriate critical event for a given situation involves taking a stand on what is the successful adaptation. As an example, assume that a car collides with a bicyclist while turning left in an intersection, and that car speed was appropriate for the turning maneuver (i.e. did not lead to skidding). Assume also that a bus was coming from the opposite direction and passed the intersection just before the car driver initiated the turn, so the bus occluded the car driver’s view of the intersecting bicycle lane. Given these facts, while Speed: surplus and Distance: shortened are certainly possible critical event codes, the one that best captures the situation would be Timing: premature action; the turning maneuver should have been performed later when the driver had a better view of the bicycle lane. Note also that, in practice, an inappropriate selection of a critical event does not create a problem downstream, since critical events which are related have similar links to the contributing factors. Hence, the choice of critical event does not limit the choice of contributing factors. Furthermore, in the SNACS, the possible links between contributing factors are predefined; the individual coder does not need to determine for each case what contributing factors link to each other. These predefined links were introduced both to preserve previously acquired knowledge on how contributing factors relate and to improve intercoder agreement (of course, a procedure exists for introducing and/or removing links as empirical knowledge increases). Consequently, whole chains of interlinked causes and consequences can be deduced; if there were only one set of direct causes, the analysis would have a great breadth but no depth. Moreover, the linking system enables SNACS charts to be easily and systematically aggregated to find common causation patterns for a particular group of charts. It should be noted that a link cannot be used just because it is given in the manual. The use of a link in an analysis must always be supported by the data available. In the SNACS, the critical event categories and the contributing factors, as well as the link system, were defined according to knowledge derived from the analyses of in-depth crash data and literature reviews (see Wallén Warner et al., 2008).

Fig. 1. An example of a SNACS chart for a driver. The SNACS links for this chart could be written as critical event-Z1–Y1, critical event-Z1–Y2–X1, and critical eventZ1–Y2–X2. This means that the critical event occurred as a result of the interaction between the Z1 and Y1 causes and the Z1, Y2, X1, and X2 causes. In this instance, Z1 is a direct cause of the critical event, a direct consequence of Y1 and Y2, and an indirect consequence of X1 and X2. On the other hand, Y2 is both a cause of Z1 and a consequence of interaction between X1 and X2. Here, Z1 is the closest contributing factor (either in time or space) to the critical event.

A1: Timing A2: Duration A3: Force/power A4: Distance A5: Speed A6: Direction A7: Object A8: Sequence

In Section 3.1 we define and discuss the most frequent critical events and contributing factors revealed by our analyses. All of these definitions are based on the SNACS manual (Ljung, 2007). It should be noted that there are events and factors in our analysis which are not defined or discussed in Section 3.1, as they were not common occurrences. However, these definitions are available in the manual (Ljung, 2007). 2.2. Macroscopic data Macroscopic data was obtained from police reported crashes in the database STRADA (Sjöö and Ungerbäck, 2007). This database contains all police reported crashes in Sweden. A total of 9702 carto-VRU intersection crashes were used in the analysis. The crashes occurred in the years 2003–2007 and involved 16,119 people, with at least one injury or fatality per crash. Cases with missing or unknown information were excluded. 2.3. Deriving driver support needs and comparing the ADAS types In this study, the causation charts retrieved were aggregated in three ways. First, all SNACS charts for the 60 drivers were aggregated to provide a complete overview of the contributing factors, and links between them. Next, the causation charts were aggregated based on the type of VRU involved (pedestrian or cyclist) and intersection type (X or T) in which the crashes occurred. This was done to assess whether different VRU and intersection types would generate causation patterns that differed from a combination of them. In each aggregation, the most frequent set of causation links was denoted as the most common causation pattern. The implication of the aggregated charts for the information and warning providing ADASs was interpreted by means of “common logic” and findings in the literature. 2.4. Derivation of sensor requirements Generally, three broad types of ADASs can be distinguished with respect to sensor configuration: infrastructure-based, vehiclebased, and cooperative systems. This study used the findings from the aggregation of contributing factors in the microscopic data analysis to determine which of these configurations is the most likely to succeed in preventing the crash type selected. In addition, to find out under what light and weather conditions an ADAS should be able to operate, we used both microscopic and macroscopic data. First, the SNACS analysis was used to investigate whether these conditions play a frequent contributing role in the car-to-VRU crashes (e.g. limited visibility due to darkness and/or bad weather). Next, we investigated the relationship between light and weather conditions and the number of injured or killed (injury severity sustained) in the car-to-VRU crashes from STRADA. This way we could, for example, determine the injury and fatality reduction potential for a sensor system limited to operating under daylight conditions. The findings from the SNACS analysis, which are based on a rela-

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580 Table 2 The relation between the posted speed limits and car speeds. Posted speed limit (km/h) 30 50 70 90 110

Car speed (km/h) 2.5%

50%

97.5%

29.3 51.0 67.3 87.9 110.2

34.5 52.4 68.4 88.9 111.4

39.7 53.8 69.5 89.9 112.6

tively limited number of cases, could also be compared with the findings from the macroscopic data. 2.5. Derivation of crash risk assessment algorithm requirements Selection of the most appropriate time for issuing a warning is a trade-off between the need for early detection and avoiding false alarms (Parasuraman and Hancock, 1999). Moreover, an early warning can be experienced by drivers as a disturbance, even if it is technically correct, either because drivers do not understand why it was issued or because they have already identified the potential problem and adapted to it. On the other hand, late warnings may not contribute to crash avoidance; they may actually disrupt an ongoing preventive action (Lee et al., 2002). Determining how different factors affect the timing of a warning in a particular pre-crash scenario is of crucial importance in ADAS development. In principle, the timing of a warning can be seen as a function of the distance needed to avoid the risk of a crash by applying a preventive action such as braking. This distance is here referred to as distance-to-avoidance (DTA), while the corresponding time is denoted as time-to-avoidance (TTA). For the estimation of these, the following equation was used: TTA =

vcar 2a

+ tdriver + tdevice

1573

conflicts in urban intersections, may be distinguished. Since only car-to-pedestrian data are available from the study, these data were also used here for car-to-bicyclist conflicts. The findings from the microscopic data analysis were used to assess which of these deceleration levels is applied in the crash type considered. As for driver reaction time (tdriver ), again this variable does not exist in the data sources used; it has to be estimated from other sources. This is potentially very complex, as the time it takes for a driver to perceive a warning and respond may be affected by demographic factors, the design of the human–machine interface (HMI), cognitive load, and situation urgency. However, according to Green, 2000, driver reaction time is mostly modulated by driver’s expectations; tdriver increases significantly for unexpected and surprise events compared to expected events. This study interpreted the expectancy level for the crash involved drivers from the SNACS charts. For example, based on the aggregated SNACS chart, we could conclude that Distraction is a frequent contributing factor. Based on “common logic” and the support from literature, we could then determine whether the distracted drivers would have expected a warning or not. These results were then used to select an appropriate tdriver from the values presented in Green, 2000. For most systems, the tdevice are much shorter and easier to obtain than tdriver . This time accounts for various delays within an ADAS, including the time needed to detect a threat, issue a warning, and to get response from the brake system after it has been deployed by the driver. The present study used a tdevice of 300 ms, as presented by Rössler et al. (2005). By estimating the TTA for a typical scenario, it is also possible to find out how much time in advance the intentions of the drivers and VRUs must be predicted. To obtain the distance that the VRUs may travel during this period of time, it is necessary to know their speeds. However, these speeds are not registered in the crash databases used here. The VRU speeds were therefore adopted from the literature.

(1)

where vcar is velocity of a car, tdevice is latencies in the actuation system, tdriver is driver reaction time, and a is deceleration. This study estimated a TTA that would apply in a typical carto-VRU pre-crash scenario. For this scenario, the parameters in Eq. (1) were selected to enable the drivers in the crashes to react to a cautionary crash warning (CCW) and an imminent crash warning (ICW), and come to a complete stop before a potential collision point. We assumed that braking is the preventive action applied by the driver (alternative actions were not considered). The approach speed of the cars in the pre-crash phase (vcar ) is not registered in the crash databases used in this study. To estimate this speed, we made a distribution of crashes according to the posted speed limit on the road for the crashes registered in STRADA. This distribution was then compared with the relationship between the posted speed limits and the car speeds presented in Svedung (2005). Svedung measured the speed of cars 100 meters prior to intersections. The measurements took place in 2004 at 1600 randomly selected intersections in Sweden. The speeds measured were then correlated to the posted speed limit, as shown in Table 2. Based on this comparison, we could conclude what was the likely approach speed of the cars in a pre-crash phase. Information on the applied deceleration (a) would ideally be found in a crash database. However, STRADA and SafetyNet databases do not contain this information as such, nor even information from which it can be deduced. For this study, therefore, data from a study by Malkhamaha et al. (2005) were used. They state that applied deceleration, relative to other conditions, is mostly affected by perceived situation urgency and may be expressed as a function of time to collision. Three levels of deceleration, which characterize potential (a1 = 0.3 g), slight (a2 = 0.5 g), and serious (a3 = 0.7 g)

2.6. Derivation of HMI requirements In addition to warning criteria, the effectiveness of a warning system will be greatly affected by the design and configuration of its warning interface, here referred to as the human–machine interface (HMI). This study focused on identifying the most suitable HMI configuration in terms of positioning. In general, an HMI can be positioned in the infrastructure, the vehicle, or both. The findings from the aggregated chart analysis were used to assess which of these HMI configurations is the most suitable for prevention of the crash type considered. 3. Results 3.1. Driver support needs 3.1.1. Causation charts for drivers in crashes involving VRUs The most frequently coded critical events in the aggregated SNACS charts were inadequate Timing and inadequate Distance. These critical event types accounted for 68% and 25% of the critical events in the 60 cases (Fig. A1). A closer investigation of the instances of inadequate Timing shows that Timing: late action was coded for 15 of 60 drivers (Fig. A1). According to the definitions in the SNACS manual, these drivers initiated a preventive action such as braking or steering. However, they started the action too late, i.e. the time available to avoid the crash was shorter than that required to complete the preventive action. For 14 of 60 drivers, Timing: no action was coded. These 14 drivers did not perform any type of preventive action prior to the crash. There were also 12 cases coded as Timing: premature

1574

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

Fig. 2. Aggregated SNACS charts with detailed information about contributing factors for car drivers involved in critical events where the critical event was triggered by Observation missed. To make the figure easier to read, we have excluded the links showing that some of these critical events occurred as an interaction between observation missed and other contributing factors. These links are shown in Fig. A2 in Appendix A1 .

action. According to the SNACS manual, this code shows that the driver started an action before it was appropriate: the crash could have been avoided if the action had been carried out later. In the crashes analyzed here, this coding has been applied to the cases in which the driver entered the road before it was clear from the pedestrians or bicyclists, and cases in which the driver started a turning maneuver too early and cut a corner where the pedestrian or bicyclist was standing or moving. A closer investigation of the instances of inadequate Distance shows that Distance: prolonged was coded for 11 of 60 drivers (Fig. A1). As explained in the SNACS manual, this code shows that a vehicle moved too far, i.e. the vehicle ended up far beyond intended position. In our analysis, this code was, for instance, chosen for a diver who did not stop at a red traffic light but had intended to do so. Distance: shortened was coded for 4 of 60 drivers. For example, some of these 4 drivers were following too close to the vehicle in front and did not see the pedestrian or bicyclist; others were positioned too close to the pedestrian or bicyclist. In this context, “too close” means that the driver did not keep a minimum distance to the other road user, which is required to safely manage the traffic situation. As shown in Fig. A1, the most frequent contributing factor immediately preceding the critical events was Observation missed (links

1 Note that SNACS allows attribution of, for example, multiple Observation missed to one driver. Therefore, the total number of links can exceed the total number of aggregated charts (i.e. the number of drivers). The numbers show the frequency of occurrence.

A1–B1, A4–B1, A5–B1, and A6–B1). This contributing factor was attributed to 48 of 60 critical events. It codes instances in which the pedestrian or the bicyclist was not noticed by the driver at the time when she/he decided to carry out the maneuver which resulted in the crash. It should be noted that the on-scene crash investigations in the SafetyNet project did not have access to eyetracking or similar devices. Hence, it is not possible to distinguish between a driver who does not look at a pedestrian at all and a driver who casts a glance at a pedestrian but without making any behavioral adaptation (often referred to as “looked but did not see”). The second most frequent contributing factor closest to the critical events was Faulty diagnosis. This factor was attributed to 14 of 60 critical events (links A1–C1, A4–C1, and A8–C1). According to the SNACS manual, Faulty diagnosis shows that the driver made an incorrect or incomplete diagnosis of the development of events prior to the crash. For example, it was coded in our analysis for the drivers who believed it is allowed to turn left in a particular intersection, while in fact it is prohibited. It was also coded for the drivers who misjudged the time that the VRU needed to reach the intersection. The third most frequent factor immediately preceding the critical events was Inadequate plan. This code was given to 7 of 60 drivers. This factor should, according to the manual, be selected when a driver makes a plan that is wrong or incomplete, i.e. does not contain all elements needed for success. As an example, some of the 7 drivers in our analysis stated that they checked only for the motorized traffic at the intersecting road, and not for the VRUs.

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

Since Observation missed was the most frequent contributing factor immediately prior to the critical event, the following discussion will focus primarily on how this particular factor relates to the other coded ones. First, Observation missed was investigated to see if it interacted with other causation factors. Here it was found that in 12 of 48 cases when the critical event was triggered by Observation missed, this happened in combination with other contributing factors (see links A1–C1, A1–D1, A1–C3, A4–C1 and A5–D1 in Fig. A2). Second, the contributing factors leading up to Observation missed were identified (as shown in Fig. 2). These factors could be grouped according to the following categories. • “Reduced visibility” due to physical obstructions, weather conditions, and/or light conditions (links B1–N4, B1–N2, B1–B1.1, B1–H5, and B1–B1.4). This group of factors was coded for 29 of 48 drivers2 . The physical obstructions were coded by three factors: Temporary obstruction to view (8 drivers), Permanent obstruction to view (3 drivers), and Permanent sight obstruction (1 driver). The first one was selected when the driver’s view was limited by another vehicle. The second one was applied when the driver’s view was obstructed by some type of vegetation, building, or container. The last one shows that the driver’s view was limited by the design of own car. The view of 9 drivers was limited by rain. The codes attributed to these drivers are Temporary obstruction to view and Other (e.g. reflections from wet road surface). The light conditions such as sun glare (Glare) and darkness (Other) were attributed to 9 instances of Observation missed. • “Reduced awareness” due to Inattention and/or Distraction (links B1–E6 and B1–E3). This group of factors was attributed to 8 of 48 drivers. The factor Inattention is in the SNACS manual defined as any condition that causes the driver to pay less attention than required for the driving task. A closer investigation of the cases analyzed here shows that this factor was selected for two drivers who completely focused on motorized traffic (Other), one driver who suddenly started coughing and could not fully pay attention to the driving (Temporary inability), one driver who was bored (Bored or unmotivated), and one driver who did not look for other road users in a particular intersection, since there are usually no other road users there (Habit or expectation). Generally, Distraction is coded when the performance of the driving task is suspended because the driver’s attention was caught by something else. It was selected if the driver was trying to enter a destination on the navigation system (Internal competing activity), the driver was fighting with a passenger (Passengers), or if the driver was paying attention to some irrelevant events close to the road (External competing activity). • “Insufficient comprehension” due to Faulty diagnosis of the situation and/or Inadequate plan of action (links B1–C1 and B1–D1). This group of factors was coded for 18 of 48 drivers. It occurred due to a variety of factors. One of the most frequent was Error in mental model (10 drivers). It reveals that the driver did not expect to see any VRUs in the intersection. The factor Psychological stress was attributed to two drivers. To exemplify, one of these drivers was too excited about newly received driving license. For two drivers, Insufficient knowledge was coded. These two drivers were inexperienced and not familiar with the car they were driving. The contributing factor Information failure between driver and traffic environment was also attributed to two drivers. According to SNACS manual, this factor shows that the driver failed to receive a key message from the surrounding traf-

2

Note that Fig. 2 shows 30 links since two factors were attributed to one driver.

1575

fic environment. In the cases analyzed this occurred due to lack of a traffic sign (Inadequate information design) and due to invisible bicycle lane (Inadequate road design). For one of the drivers Overload or too high demands was coded. This was due to insufficient management of traffic flow which placed high demands on drivers. 3.1.2. Comparison of SNACS charts for drivers involved in VRU crashes in X and T intersections This study considers two intersection types, X and T. When SNACS charts are aggregated separately for these two types, the causation patterns are in general similar to the aggregated SNACS chart combining both intersection types. However, a comparison of the aggregated charts for X and T intersection reveals some (minor) differences between them. The critical event Distance was more frequently coded in X intersections (29%) than in the T intersections (20%). Also, “insufficient comprehension” was more frequently coded for the drivers in X intersections (77%) than for the drivers in T intersections (48%). “Reduced visibility” was more frequent in X intersections (43%) than in T intersections (32%). 3.1.3. Comparison of SNACS charts for drivers in crashes with pedestrians or with bicyclists When SNACS charts are aggregated based on the type of VRU involved (pedestrians or bicyclists), the causation patterns identified are very similar to the patterns in the aggregated SNACS charts for both types of road users combined. However, some differences were identified in the frequency of the following events and factors: • Critical events involving inadequate Timing were more frequent in car-to-pedestrian crashes (81%) than in car-to-bicyclist crashes (62%); • Critical events involving inadequate Distance were more frequent in car-to-bicyclist crashes (31%) than in car-to-pedestrian crashes (14%); • Observation missed was more frequently coded for the drivers in car-to-pedestrian crashes (90%) than for the drivers in car-tobicyclist crashes (74%); • “Insufficient comprehension” was coded more frequently for the drivers in the bicyclist crashes (74%) than for the drivers in the pedestrian crashes (57%); and • “Reduced visibility” was coded more frequently for the drivers in the pedestrian crashes (52%) than for those in the crashes involving bicyclists (31%). 3.2. Comparison of information providing and warning providing ADAS Based on the results from aggregated charts for the drivers in car-to-VRU intersection crashes, some tentative conclusions about the first two ADAS categories presented by Carsten and Nilsson (2001) can be drawn. Information ADASs aim to help drivers to navigate and informs them about the road conditions, traffic environment, and vehicle state (Carsten and Nilsson, 2001). In relation to VRU crashes, a system which informs drivers that they are about to face a potentially hazardous situation is technologically quite feasible. For example, all intersections in the map database of the navigation system could be tagged according to how often car-toVRU crashes have occurred in them before; the driver could be informed about intersections above a certain frequency threshold. However, the most frequent contributing factors identified in the aggregated causation charts indicate that the potential benefit of a system providing only general crash risk information may be limited. There are several reasons for this. First, since many of the crashes were caused by “reduced visibility” (29 of 48, see Fig. 2), drivers may be unable to see the VRUs despite any information

1576

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

Table 3 Fatality and injury distribution under various weather, light, and road conditions based on STRADA. Cases unknown conditions are excluded. Weather conditions

Fatalities and injuries

Light conditions

Fatalities and injuries

Dry Raining, fog, or snowing

6193 (86%) 974 (14%)

Daylight Darkness, dusk, or dawn

5491 (74%) 1907 (26%)

provided on general risk. Second, as many drivers (8 of 48) were coded as Inattentive and Distracted, the information provided by such an ADAS may not be noticed. The latter is supported by several studies which show that information about potentially hazardous situations is often disregarded by drivers, as crashes are rare events and drivers consider their own involvement in crashes unlikely (Jonsson et al., 2008). Moreover, providing information on general risk may divert drivers’ attention from the main driving task in an already demanding situation, and the amount of information provided could exceed either what drivers can handle or what they will tolerate. For a particular group of crashes, a general information providing ADAS may be suitable, namely crash events when missed observations occur due to faulty diagnosis of the situation; this, in turn, is caused by information failure between drivers and the traffic environment (2 of 48, link B1-C1-J2 in Fig. 2). A closer reading of the cases in the SafetyNet database reveals that these crashes include events which occurred, for instance, due to missing traffic regulation signs and/or inappropriate placement of a bicycle lane. Such crashes may be avoided by providing drivers with information about traffic regulations and intersection geometry. However, rather than tagging previous crash statistics onto a map database, the system’s decision to provide information should be based on a real-time calculation of crash risk, which is guided by dynamic data from the traffic environment, in order to minimize the problems of increased cognitive load and/or reduced driver acceptance. For warning ADASs, two types can be distinguished: imminent crash warning (ICW) and cautionary crash warning systems (CCW), as described by Campbell et al. (2007). An ICW is generally used in situations where immediate corrective action by the driver (such as braking or steering) is required. In particular, an ICW is viewed as useful when drivers are distracted, inattentive, or when sudden changes in the traffic environment occur, such as animals running out on the road. Since a large proportion of the drivers in car-toVRU crashes analyzed in Section 3.1 were coded as having “reduced visibility” (29 of 48) and/or “reduced awareness” (8 of 48) as contributing factors, an ICW system would most likely help. On the other hand, since many occurrences of Observation missed were attributed to “insufficient comprehension” of the situation (18 of 48), there is also a need for the earlier type of warning given by a CCW system. The CCW may help drivers reconfigure their perception of how events will unfold and adjust their driving plans accordingly, thus eliminating hard braking and other undesirable actions which may be caused by startle effects from ICWs. However, since CCWs generally are issued more often, earlier, and contain more detailed information about upcoming situations than ICWs, false alarm rates may go up. This could lead to reduced reliance on, and use of, the system. Also, although detailed warnings allow a better understanding of the situation, they may be more demanding to comprehend if presented with inappropriate formatting and timing. Given that avoiding car-to-VRU crashes may require both ICW and CCW, a combination designed to overcome the weaknesses of each warning type would seem to be a suitable way forward. Another way to reduce the risk of car-to-VRU crashes at intersections would be to engage intervention systems to partly or completely relieve drivers from control of the vehicles. This would correspond to the two intervention categories defined by Carsten and Nilsson (2001). However, while the application of such systems in principle could be highly beneficial, the microscopic data

used in this study contains little information that would be helpful in deriving requirements for such systems, at least in terms of the kinematic details and information on drivers’ reactions to interventions. A closer reading of the SNACS coding manual (Ljung, 2007) reveals that this actually seems to be intentional. According to Girard (1993), crashes can be divided into four phases: (1) the driving phase (the “normal” driving situation where no unexpected demands are made on the driver), (2) the discontinuity phase (the “normal” driving situation is interrupted by an unexpected event), (3) the emergency phase (the time and space between discontinuity and impact) and, finally, (4) the crash phase (the crash and its consequences). According to the SNACS coding manual, the SNACS analysis is deliberately focused on contributing factors which lead to the discontinuity phase, while details on the emergency phase (i.e. where intervention systems would be active) are left to other projects to investigate. The macroscopic crash data also lacks the type of detail which would aid intervention requirement specification. Taken together, these findings indicate that there is a gap in currently collected crash data. Many field operationl tests are currently underway (e.g. SHRP2, 2007; SeMiFOT, 2009), which may help to bridge this gap. What can be concluded for now, however, (i.e. based on the aggregated SNACS chart analysis in Section 3.1), is that a combination of CCW and ICW systems, with the added capability of providing information on traffic regulations and intersection outline in relevant situations, has a great potential to reduce the risk of car-to-VRU crashes in urban intersections. On a final note, the separate causation patterns for the two intersection types and the two VRU types did not deviate significantly from the combined patterns. Consequently, the conclusions drawn here would hold even if only one type of intersection or VRU were studied. 3.3. Functional requirements 3.3.1. Sensor requirements In the aggregated SNACS charts analysis, it was concluded that the risk of car-to-VRU crashes at intersections may be significantly reduced by helping drivers to notice VRUs. This means that a safety system should be able to address the causes of Observation missed. The majority of these misses did not occur in conjunction with unsatisfactory weather or light conditions. However, for 18 of 48 drivers who failed to observe the VRUs, unsatisfactory weather and light conditions such as rain, glare, and darkness were coded as contributing factors (Fig. 2). This is corroborated by the macroscopic data analysis. It shows that, although 70% of fatalities and injuries occurred under good weather and light conditions, 30% were injured or killed under adverse conditions (Table 3). Given these facts, it can be concluded that more than 70% of the crashes could be addressed by a system operating under daylight and dry weather conditions. Consequently, sensors should be able to operate when scene illumination is about 10,000–25,000 lux (daylight) and atmospheric transmission is about 0.5–4 dB/km (dry weather). However, to address all crashes analyzed, the illumination range is between 30 lux (poorly-lit street) and 100,000 lux (direct sunlight). Similarly, the required atmospheric transmission level is about 70 dB/km (dense fog). The analysis of causation factors showed that 25% of Observation missed (or 20% of 60 critical events) was, in addition to “reduced

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580 Table 4 Distribution of car-to-VRU crashes that occurred at intersections in Sweden, between 2003 and 2007, by posted speed limit. Speed limits (km/h)

Number of accidents (% of total)

30 50 70 Other

614 (8%) 6680 (84%) 471 (6%) 135 (2%)

awareness” and “insufficient comprehension”, caused by physical obstructions of the drivers’ view. These obstructions included vegetation, other vehicles, buildings, and host car design (Fig. 2). This indicates that an ADAS targeting these crashes should be configured to assist drivers when their view is obstructed. Consequently, sensors incorporated in such a system should be able to detect objects in “blind-spots” that cannot be seen by the drivers. Although vehicle-based systems provide sufficient detection of VRUs on the road, they usually have a limited field of view (Gandhi and Trivedi, 2007). To overcome this, vehicle-to-vehicle or vehicleto-infrastructure communication needs to be used. In particular, by positioning sensors in the infrastructure and sending dynamic data about objects detected in real-time to approaching vehicles, drivers could be warned of collision risk with VRUs who are impossible to detect with vehicle-based sensors. 3.3.2. Crash risk estimation requirements The aggregated SNACS chart analysis (see Section 3.1), showed that 48 of 60 (or 80%) drivers did not notice the VRUs before they collided. Hence, there is reason to believe that a warning would have been perceived by these drivers as a surprise event. For surprise events, Green, 2000 approximates drivers’ brake reaction time (tdriver ) to 1.5 s. Since 80% of the drivers in our SNACS analysis is likely to have a tdriver of 1.5 s, this is the time that we used for estimation of TTA. As this is an average value, both longer and shorter tdriver may be experienced in real traffic situations. In addition, presenting additional information with warnings, such as intersection outline and traffic regulations, may result in an increased tdriver . For the estimation of vcar , the distribution of the car-to-VRU crashes was made by posted road speed limit (see Table 4). As Table 4 shows, 7294 of 7900 crashes occurred in the intersections with a posted speed limit of 50 km/h, or lower. Based on this distribution and the relationship presented in Table 2, it is likely that the average car speed, prior to 92% of the crashes, was 52 km/h, or lower. Table 2 reveals that the car speeds prior to the crashes that occurred in intersections with a posted speed limit of over 50 km/h exceeded 67 km/h. However, the number of these crashes is 606 which is much lower than the number of the crashes that occurred on the roads with the posted speed limits 50 km/h, or lower. Therefore, we have concluded that vcar prior to the majority of car-to-VRU crashes was about 52 km/h. By inserting the values for tdriver and vcar and the values presented in Section 2.5 in Eq. (1), the time-to-avoidance (TTA) and the corresponding distance-to-avoidance (DTA) were estimated. Based on these estimations, the following conclusions can be drawn. • The minimum DTA and TTA are about 40 m and 2.8 s, respectively, assuming that the warning is a surprise for the driver and deceleration is the same as in a serious conflict (a3 ). Since ICWs have the function to initiate immediate braking by the driver and are usually issued close to a potential collision point, it can be assumed that drivers in these situations decelerate as if in a serious conflict. This minimum distance and time correspond therefore to the point in space and time where an ICW should be issued. To avoid crashes with a system providing CCW, the corresponding

1577

values are about 46 m and 3.2 s, assuming a deceleration as in a slight conflict (a2 ). The TTA may increase for a system providing warnings and information on traffic regulations and intersection outline. • Both drivers and VRUs may change their paths during each of these periods of time. In a typical scenario, the warning systems mentioned will have to predict driver and VRU intentions at least 2.8 and 3.2 s in advance, respectively. During these time periods, a pedestrian walking at a normal speed of 1.5 m/s may travel at least 4.2 and 4.8 m, respectively. Corresponding distances traveled for a bicyclist would be about 11.2 and 12.8 m, assuming a speed of 4 m/s. This shows that each of the warning systems will need to accurately “know” the position and velocity of the VRUs in order to predict their future trajectories. Also, it is likely that classification of the VRUs will be required since, for example, the choice of crossing place may differ for pedestrians and bicyclists. The ability to track the VRUs for multiple signal samples will also be necessary to understand the traffic situation and thereby reduce the number of false alarms.

3.3.3. HMI positioning requirements The analysis of the aggregated SNACS charts for the drivers involved in car-to-VRU crashes indicates that an HMI positioned in the infrastructure may not be as beneficial as an HMI positioned in the vehicle. As Section 3.1 shows, “reduced awareness” was attributed to 16 of 48 drivers who failed to see the VRU. It may not be possible to redirect the attention of these 16 drivers by a warning issued in the infrastructure. For example, a recent study by Inman and Davis (2009) investigated the potential effectiveness of in-vehicle and infrastructure based warnings to prevent red-light violations by vehicle drivers. The study showed that the infrastructure based warnings are less effective than the in-vehicle warnings. In addition, several studies show that auditory warnings may be especially useful for drivers with “reduced awareness” (Campbell et al., 2007; Lerner et al., 1996). However, issuing such warnings from the infrastructure may be difficult, especially when it comes to urban traffic environments. Also, it may be difficult for drivers to perceive who is the target of an infrastructure-based warning. Given these facts, it can be concluded that an HMI should preferably be positioned in the vehicle.

4. Discussion Although a set of active safety system requirements could be derived based on a thorough understanding of real-world crashes, the method applied here has limitations. One is the limited number of cases in the aggregation, which implies that results are not statistically representative. On the other hand, in-depth multidisciplinary investigations provide information unattainable in other ways (Sandin, 2008), and the data has been collected from a wide geographical area, which means the causation patterns identified are widespread. This is based on the assumption that people experience similar problems, irrespective of geographical and cultural differences, as long as they are performing similar activities under similar circumstances. In addition, the in-depth cases analyzed cover a range of vehicle and VRU trajectories prior to crashes. This study did not aggregate the causation charts according to the trajectory type. However, one could expect different causation patterns depending on the precise vehicle and VRU trajectories prior to crashes. A natural step in further research would be to aggregate charts according to such a criterion, and to compare those causation patterns with the present study.

1578

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

This study aggregated the causation charts for drivers involved in crashes with at least one injury or fatality. This approach was taken with the Swedish Vision Zero concept (Johansson, 2009) in mind. According to this concept, no one should be killed or injured within the road transport system. Consequently, it is equally important to address all crashes independently of the severity level. However, one could expect different causation patterns, and ADAS requirements, depending on the crash severity. A direction of further research should therefore be to aggregate charts according to the injury severity of the crash and to investigate the implications for designing ADASs. The study aggregated and analyzed causation charts for the drivers in 60 crashes. In connection with that, it also aggregated the causation charts for the VRUs in these crashes. The results indicated that further understanding of these causes may aid the design of ADASs. For instance, it may help to specify algorithms for the prediction of VRU behavior prior to crashes. It may also help researchers to identify situations in which a pedestrian assistance system could be more effective than a driver assistance system. As these issues were not addressed here, this could be a topic for further research. An additional limitation of the present study concerns the method applied to transform the knowledge, generated by establishing causation patterns, into requirements. The SNACS only provides systematic classification of the contributing factors; it does not provide any method to interpret these factors for the development of systems. While “common logic” can be applied with satisfactory results, as it has been in this study, the study highlights a great need for such a method in order to achieve a higher level of consistency. The study also recognizes the importance of other data sources for the design of ADASs. First, the study did not examine the potential of intervention systems. The crash data used lack the type of detail which would aid the specification of such systems. Second, to determine the TTA various parameters were approximated due to a lack of precise data. For example, the car speed was estimated based on the speed measurements 100 m prior to intersections (Svedung, 2005). However, it is possible that some drivers slow down, or accelerate, closer to an intersection. Consequently, the TTA may be overestimated for some drivers, and underestimated for some others. It is likely that data collected in naturalistic driving studies or field operational tests will address these problems. From successful projects of this kind, several data sets that are unattainable from crash investigations could be obtained. On a final note, the study may have overestimated, or underestimated, the TTA due to the assumption that a warning will result in a complete stop. In real traffic, drivers usually adapt their speed to let the VRUs cross the road without coming to a stop. A speed reduction of, for instance, 90% would give a significantly shorter TTA. On the other hand, a complete stop may be necessary to avoid some crashes. Also, the study assumed that the only preventive

action is braking. This limitation was chosen since braking, compared with other preventive actions, requires the longest time to perform (Green, 2000). However, the potential of other preventive actions should be explored, especially in connection with the assessment of intervening systems. 5. Conclusions Some preliminary requirements that an advanced driver assistance system (ADAS) should fulfill to reduce the number of car-to-VRU crashes at intersections were derived in this study. The derivation was based on the aggregation of individual causation charts for drivers involved in car-to-VRU crashes, analysis of macroscopic crash data, and literature studies. The requirements identified can be summarized as follows. • The primary aim of an ADAS should be to help drivers to observe the VRUs in time. This includes enhancing the drivers’ ability to interpret the development of events in the near future. • This ADAS should provide a combination of imminent and cautionary crash warnings, with additional support in the form of information about intersection geometry and traffic regulations where relevant. • These warnings and information should be deployed by means of an in-vehicle HMI, and according to the likelihood of crash risk. To predict the crash risk accurately, the road users should be tracked over time and their intended path predicted. It may be necessary to predict the paths 3.2 s, or more, ahead. This means that a pedestrian’s intention to cross the road should be predicted approximately five meters ahead. The corresponding value for bicyclists is 13 m. • The ADAS operating under good weather and light conditions are expected to prevent about 70% of the car-to-VRU crashes. To prevent as many crashes as possible, it must also operate under adverse conditions. It should have the capacity to support drivers when their view is obstructed by physical objects, such as buildings and other vehicles. To address problems that may arise for vehicle-based sensors, the use of cooperative systems based on infrastructure-to-vehicle communications is recommended. Acknowledgments This research was conducted as a part of the VISAS project. It was financially supported by the Swedish Governmental Agency for Innovation Systems (VINNOVA), Autoliv, Volvo Car Corporation, and Chalmers University of Technology. The authors thank the European project SafetyNet for providing the crash data. Helen Fagerlind and Mikael Ljung Aust at Chalmers and Erik Coelingh at Volvo Car Corporation are thanked for their valuable comments on earlier drafts of the manuscript.

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

1579

Appendix A.

Fig. A1. Aggregated SNACS charts for 60 car drivers in crashes involving pedestrians and bicyclists in X and T intersections. Of these drivers approximately 35% were involved in crashes with pedestrians and 65% in crashes with bicyclists. The crashes occurred in urban areas (speed limit <60 km/h). The numbers indicate the frequency of causation factors and link. Note that SNACS allows the coding of, for example, multiple Observation missed for one driver. Therefore, the total number of links exceeds the total number of drivers.

Fig. A2. Aggregated SNACS charts for 48 car drivers involved in critical events that were triggered by a combination of Observation missed and other contributing factors.

1580

A. Habibovic, J. Davidsson / Accident Analysis and Prevention 43 (2011) 1570–1580

References Andreone, L., Bekiaris, E., Guarise, A., Mousadakou, A., 2006. D2.1 – Requirements and use cases. Report no.: D2.1. Vehicle-to-Vulnerable roAd user cooperaTive communication and sensing teCHnologies to imprOVE transpoRt safety (WATCH-OVER). Retrieved April 15, 2009, from www.watchover-eu.org. Björkman, K., Fagerlind, H., Ljung Aust, M., Liljegren, E., Morris, A., Talbot, R., Danton, R., Giustiniani, G., Usami, D., Parkkari, K., Jaensch, M., Verschragen, E., 2008. SafetyNet Deliverable 5.8: In-depth accident causation database and analysis report. European Commission, Directorate-General Transport and Energy. Retrieved April 15, 2009, from http://www.erso.eu. Campbell, J.L., Richard, C.M., Brown, J.L., McCallum, M., 2007. Crash Warning System Interfaces: Human Factors Insights and Lessons Learned. Report, no.: DOT HS 810 697, National Highway Traffic Safety Administration (NHTSA), Washington, D.C. Carsten, O.M., Nilsson, L., 2001. Safety assessment of driver assistance systems. European Journal of Transport and Infrastructure Research (EJTIR) 1 (3), 225–243. ERSO, 2008. Annual Statistical Report 2008; Based on CARE/EC. European Road Safety Observatory (ERSO), Retrieved April 15, 2009, from http://www.erso.eu. Gandhi, T., Trivedi, M.M., 2007. Pedestrian protection systems: issues, survey, and challenges. IEEE Transactions on Intelligent Transportation Systems 8 (3), 413–430. Gavrila, D., Marchal, P., Meinecke, M.-M., 2003. Deliverable 1-A: vulnerable road user scenario analysis. Sensors and system architecture for vulnerable road users protection (SAVE-U). Girard, Y., 1993. In-depth investigation of accidents: the experience of INRETS at Salon-de-Province. In: International Congress on Safety Evaluation of Traffic Systems, Traffic Conflicts and Other Measures, Salzburg. Green, M., 2000. How long does it take to stop? Methodological analysis of driver perception-brake times. Transportations Human Factors 2 (3), 195–216. Hollnagel, E., 1998. Cognitive Reliability and Error Analysis Method (CREAM). Elsevier Science Ltd, Oxford. Inman, V.W., Davis, G.W., 2009. Effects of In-Vehicle and Infrastructure Based Collision Warnings to Non-violating Drivers at Signalized Intersections. Report no.: FHWA-HRT-09-049, U.S. Department of Transportation, Federal Highway Administration (FHWA). Retrieved September 27, 2010, from http://www.fhwa.dot.gov/publications/. Johansson, R., 2009. Vision zero – Implementing a policy for traffic safety. Safety Science 47 (6), 826–831. Jonsson, I.-M., Harris, H., Nass, C., 2008. How accurate must an in-car information system be?: Consequences of accurate and inaccurate information in cars. In: Proceeding of the Twenty-Sixth annual SIGCHI Conference on Human Factors in Computing Systems, Florence, pp. 1665–1674. Lee, J.D., McGehee, D.V., Brown, T.L., Reyes, M.L., 2002. Collision warning timing, driver distraction, and driver response to imminent rear-end collisions in a highfidelity driving simulator. Human Factors: The Journal of the Human Factors and Ergonomics Society 44 (2), 314–334. Lerner, N.D., Dekker, D.K., Steinberg, G.V., Huey, R.W., 1996. Inappropriate alarm rates and driver annoyance. Report no.: DOT HS 808 533, National Highway Traffic Safety Administration (NHTSA), Washington, DC. Ljung, M., 2002. DREAM – Driving Reliability and Error Analysis Method. Master Thesis, ISRN: LIU-KOGVET-D-02/16-SE, Department of Computer and Information Science (IDA), Linköping University, Linköping, Sweden. Ljung, M., 2007. Manual for SNACS: SafetyNet Accident Causation System, version 1.2. Chalmers University of Technology, Gothenburg. Ljung Aust, M., 2010. Generalization of case studies in road traffic when defining pre-crash scenarios for active safety function evaluation. Accident Analysis and Prevention 42 (4), 1172–1183.

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. Lundberg, J., Rollenhagen, C., Hollnagel, E., 2009. What-you-look-for-is-what-youfind – The consequences of underlying accident models in eight accident investigation manuals. Safety Science 47 (10), 1297–1311. Malkhamaha, S., Tight, M., Montgomery, F., 2005. The development of an automatic method of safety monitoring at Pelican crossings. Accident Analysis and Prevention 37 (5), 938–946. NHTSA, 2008. Motor Vehicle Traffic Crash Fatality Counts and Estimates of People Injured for 2007. Report no.: DOT HS 811 034, National Center for Statistics & Analysis, National Highway Traffic Safety Administration (NHTSA), Washington, D.C. Parasuraman, R., Hancock, P.A., 1999. Using Signal Detection Theory and Bayesian Analysis to Design Parameters for Automated Warning Systems. In: Scerbo, M., Mouloua, M. (Eds.), Automation, Technology, and Human Performance: Current Research and Trends. Lawrence Erlbaum Associates, Mahwah, NJ, pp. 63–67. Reason, J., 1997. Managing the Risks of Organizational Accidents. Ashgate Publishing, Burlington, VT. Rössler, B., Obojski, M.A., Stanzel, M., Ehmanns, D., Kosch, T., Hopstock, M., Shooter, C., Heenan, A., Woolley, P., Benson, C., Boissou, J.-F., Martail, Y., Serezat, L., Page, Y., Chauvel, C., Fürstenberg, K., Mousain, J., de La Fortelle, A., Parent, M., Deutschle, S., 2005. Requirements for intersection safety applications. Sub-project Deliverable D40.4 of the INTERSAFE project, INTERSAFE Consortium. Salmon, P.M., Lenné, M.G., Stanton, N.A., Jenkins, D.P., Walker, G.H., 2010. Managing error on the open road: the contribution of human error models and methods. Safety Science 48 (10), 1225–1235. Sandin, J., 2008. Aggregating Case Studies of Vehicle Crashes by Means of Causation Charts. Ph.D. Thesis, Department of Applied Mechanics, Chalmers University of Technology, Gothenburg, Sweden. Sandin, J., 2009. An analysis of common patterns in aggregated causation charts from intersection crashes. Accident Analysis and Prevention 41 (3), 624–632. SeMiFOT, 2009. Sweden Michigan Naturalistic Field Operational Test. Retrieved November 14, 2009, from www.chalmers.se/safer. SHRP2, 2007. Asking Why to Learn How – 2007 Annual Report of the Second Strategic Highway Research Program. Retrieved November 14, 2009, from www.trb.org/SHRP2. SIKA, 2008. Road Traffic Injuries 2007. Report: SIKA statistik 2008:27. Swedish Institute for Transport and Communications Analysis (SIKA), Östersund, Sweden. Retrieved April 15, 2009, from www.sika-institute.se. Singh, S.K., 2005. Review of urban transportation in India. Journal of Public Transportation 8 (1), 79–97. Sjöö, B., Ungerbäck, A.-C., 2007. Nytt nationellt informationssystem för skador och olyckor inom hela vägtransportsystemet: STRADA (New national information system for injuries and accidents within the entire road transportation system: STRADA). Report no. 2007:147. Swedish Road Administration, Borlänge, Sweden. Svedung A.C., 2005. Speeds and Time Gaps (in Swedish: Hastigheter och tidluckor). Report no. 2005:2. Swedish Road Administration, Sweden. 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. Wallén Warner, H., Björklund, G., Johansson, E., Ljung Aust, M., Sandin, J., 2008. DREAM 3.0: Documentation of references supporting the links in the classification scheme Report. Chalmers University of Technology, Göteborg.