Proximity hazard indicator for workers-on-foot near miss interactions with construction equipment and geo-referenced hazard areas

Automation in Construction 60 (2015) 58–73

Contents lists available at ScienceDirect

Automation in Construction journal homepage: www.elsevier.com/locate/autcon

Proximity hazard indicator for workers-on-foot near miss interactions with construction equipment and geo-referenced hazard areas Jochen Teizer ⁎, Tao Cheng RAPIDS Construction Safety and Technology Laboratory, Ettlingen, Germany

a r t i c l e

i n f o

Article history: Received 29 November 2014 Received in revised form 3 August 2015 Accepted 20 September 2015 Available online xxxx Editor: M.J. Skibniewski Keywords: Construction workers-on-foot and equipment blind spots Behavior-based safety Automated data fusion and mining Occupational safety and health design and planning auditing, education, and training Lagging and leading safety indicators Near miss and proximity hazard events Real-time resource location tracking

a b s t r a c t Despite the many existing best practices in safety, the construction industry lacks automated safety monitoring and analysis of task-level construction operations. Data to workforce, equipment, and the overall site safety performance are currently observed, measured, and evaluated almost always manually. Such resulting performance information is likely assessed infrequently and due to subjective human interpretation or error. Research in lagging and leading safety indicators shows further that safety knowledge is hardly ever shared among relevant project stakeholders in time to prevent accidents. Since a large number of all construction fatalities are related to struck-by events – for example, workers-on-foot being too close to construction equipment and to other restricted or geo-referenced hazard areas – a novel framework around real-time location tracking technology was designed and tested to collect and study near miss data. The objectives of this article are to automatically identify the areas of static and dynamic hazards on a construction site and to automatically gather and analyze the spatial–temporal conflicts between workers-on-foot and the identified hazards. Automated conversion of raw sensor location data collected to the operation of workers, equipment, and geo-referenced hazard areas into meaningful proximity-related safety information is introduced. Field experiments validate the research based on an a priori created safe site layout information model. Results are in particular useful for practitioners or researchers who would like to enhance their quantitative and visual understanding of operational construction resource activity monitoring and analysis, and in the specific domain of detecting and mapping spatial–temporal proximity relationships of near miss events. Applications of the resulting knowledge are explained in the context of empowering construction safety engineers, managers, and the workforce by enhancing decision making in safe site layout design and planning and providing additional interactive tools in safety education and training. © 2015 Elsevier B.V. All rights reserved.

1. Introduction Construction sites have unique sizes and settings, but generally are composed of similar types of resources, for example, workers, equipment, and materials. In order to meet construction schedules, highly complex and dynamic construction activities require workers to frequently be in close proximity to potentially hazardous site conditions. Although design for safety (DfS) concepts [1] propose to eliminate safety hazards early in a project's lifecycle, construction planning inherently includes unsafe work conditions that either were overlooked by designers or planners [2] or occur due to late or frequent design change and the dynamic nature of the construction site. Examples of such hazards are poor site layout plans, insufficient site traffic control, frequent heavy equipment operations nearby workerson-foot, restricted entrance into hazardous or confined areas, or access to uncontrolled hazardous substances [3].

⁎ Corresponding author. Tel.: +49 157 54769592. E-mail address: [email protected] (J. Teizer).

http://dx.doi.org/10.1016/j.autcon.2015.09.003 0926-5805/© 2015 Elsevier B.V. All rights reserved.

Statistics show that working in proximity to hazards significantly contributed toward the number of construction fatalities. In between the years 2003 and 2010, 3171 workers were killed due to the exposure to various hazardous situations, including (1) contact with objects and equipment, (2) falls from floors to lower levels, (3) exposure to chemicals and flammable substances, and (4) struck by vehicles [4,5]. This large number of fatalities accounted for approximately 40% of all construction fatalities in those years. About one quarter of these relate to struck-by events [4,5]. Research also found the risk factors which cause worker exposure to hazardous situations. Examples are constantly changing construction site environment and conditions, unskilled laborers, high diversity of work activities occurring simultaneously, and exposure to hazards resulting from own work as well as from activities in the same or nearby locations [6]. According to these risk factors, hazards can be categorized as chemical, physical, biological, and ergonomic. Alternatively, this research classifies a hazardous situation into either static or dynamic based on the spatial–temporal characteristics of the hazard. Frequent hazardous situations occur when dynamic resources, such as heavy construction equipment, operate in close proximity to

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

workers-on-foot. Such conditions exist frequently in a congested work environment, for example, during conventional excavation in tunnels and for a high number of workforce at large capital facility projects. The U.S. Bureau of Labor Statistics reported in 2009 that among the 818 construction-related fatalities in the US, 18% (151 fatalities) were caused by workers being struck by an object or a piece of construction equipment [7]. Past research has also shown that being struck by equipment increases significantly the severity of injuries and/or the potential for a fatality of the construction personnel [8–11]. Other types mentioned – which are predominantly static in nature – are hazards such as toxic, chemical, and flammable substances, highvoltage power lines, edges in elevation, and blind spaces, for example, of ground vehicle or crane operators [12,13]. Toxic and chemical substances include dusts, mixtures, and common materials such as paints, fuels, and solvents [14,15]. High-voltage power lines pose hazards to the safe operation of cranes and derricks [12]. Falls from floor openings and leading edges have been a major reason of construction fatalities for the past years [15]. [11,12] and [16–24] stated that equipment operator visibility and specifically operator blind spaces contribute to contact collisions between equipment and workers-on-foot. Hazard controls following the Occupational Safety and Health Administration's (OSHA) safety rules and regulations and company– individual administrative policies and best practices are vital in preventing proximity-related equipment and worker incidents. Although they have been successfully established and practiced on construction projects for many years, specific understanding, evaluation, and assessment of the interaction between equipment and workers-on-foot have been missing. Furthermore, changing a company's safety culture – which is the likely approach of many safety professionals to achieve a better operational safety performance – highly depends on reliable access to accurate data. Should the problem of struck-by incidents between workers-on-foot and construction equipment in the construction industry be solved, first a scientific method is needed that can study and analyze the spatial–temporal relationship. Since humans perform existing safety data collection and analysis manually [25], the nature of resulting safety measurement is subjective and varies considerably from inspector to inspector [26]. Concepts focusing on virtual fencing of hazardous areas [27,28] and representing workspaces of equipment [29] have been discussed in several past research articles. The problems of automating the data gathering and processing of proximity events between workers-on-foot and equipment and proving that such methods promise feasible approaches to reduce incidents in the harsh construction environment have yet to be studied in greater detail. A need still exists for methods at the construction stage that measure the construction safety performances in an objective, consistent, and reliable manner. Accurate and emerging remote sensing technology and data mining algorithms can provide information from such data that are critical in the specific spatial–temporal analysis between construction workers-on-foot and equipment. Such automated methods have high potential to lead to much needed change in construction safety engineering and management practices since more detail to site resource maneuvers and behavior becomes available. Once the data are processed, the resulting information can be used for designing or planning out the hazard in the first place, or for controlling safety more effectively and efficiently during the construction processes in (near) real time. 2. Background on safety data collection and processing methods There are a variety of safety performance measures that are in use in the construction industry. The most common ones can be categorized into lagging and leading indicators. Lagging indicators are based on fatality and injury statistics. Examples include lost workday/restricted work activity injuries, and injuries

59

recorded by OSHA. Since the reporting of lagging indicators has been standardized, they are good for benchmarking the own records against others (national or competitor's safety performance). However, such reporting or benchmarking is often voluntary. Many industries apply lagging indicator data reporting as they show trends of past safety performance [30]. Studies have been conducted to demonstrate the effectiveness, efficiency, and reliability of various lagging indicators [31–33]. However, the main disadvantage of lagging indicators, especially in reflecting the safety performance in complex and dynamic construction projects, is it requires an occurrence of a reportable incident in order to count a data point [26]. Lagging indicator data thus can neither be used to prevent the occurrence of an incident, nor can it reflect the potential severity of an event, merely the consequence [34]. On the other hand, leading indicators represent a continuous monitoring of safety in ongoing work processes. Leading indicator data primarily and often focuses at the level and the analysis of small units (e.g., behavior of individuals). Hence, modifications or improvements to existing processes and behavior can be made before an incident actually occurs on a construction site [35]. Leading indicators might as well predict the future safety performance based on selected criteria [35]. Behavior-based safety (BBS), as an example of a leading safety indicator, is the application of behavioral research on human performance to the problems of safety in the workplace [36]. This technique relies on manual site observations and individual feedback after the observation period ends. The data gathered from the observation(s) is matched to a pre-defined checklist. Eventually unsafe trends can be flagged and used for pro-active resolution. Changes to safety engineering and management can be taken as needed to prevent such identified hazards in the first place. Multiple researchers have come up with a reporting scheme that is generated for gathering, analysis, control, and use of leading indicator data [37–42]. Other techniques exist to capture leading indicator data. Safety audits, for example, attempt to assess the safety management and safety culture by measuring whether selected safety performance indicators are present or not [43]. This technique is useful to gauge the extent to which an organization's policies and rules are being followed and how they might be improved. However, the effectiveness of a safety audit can be influenced by the organization's safety culture itself [44]. Investigation of near miss occurrences is another very useful measure of health and safety performance at a project level which enables organizations to learn from such errors [45]. A common industry problem is that the accuracy of reporting, counting, and analysis of near misses largely depends on a voluntary report by workers, supervisors, and management. Often, near misses are not reported due to the lack of a definition what a near miss incident is, the infrastructure reporting an incident, and motivation of organizations sharing feedback with the person that reported the near miss in the first place. In most of the leading indicator techniques, the data collection process relies heavily on manual reports [25] which cause the safety measurements to be error-prone and subjective, and ultimately considerably incoherent from inspectors [26]. Therefore, a need exists for methods that can measure specific construction safety performances in an objective, consistent, and reliable manner, and preferably in (near) real-time as on-site hazards, which were not mitigated at the design and construction planning phase, might still exist at the operational level (in the field). Emerging technologies recently have shown to be well-suited in gathering detailed site-specific safety data to construction resources. Since the proposed approach focuses on spatial–temporal proximityrelated issues between project resources (workers, equipment, and other static hazards), the following review highlights only potential technologies which are capable of providing real-time location and timestamps to a vast number of resources in the dynamic and complex nature of a construction project. [46–48] have already reviewed criteria and studied suitability of potential technologies that allow real-time location tracking of resources

60

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

at large-scale construction operations today. They concluded that realtime resource location tracking using wireless technologies is practical (low position measurement errors while the workforce accepts using the technology) and once corporate decision makers tolerate the capital investment in sensing technology. Some of the other technologies reviewed are promising, but are either technologically not as advanced or provide limitations for deployment on workforce and in safety applications [49]. For example, (a) vision-based tracking algorithms used to analyze video footage are computationally very intensive and to date do not provide reliable real-time approaches that work in object-cluttered environments like they are very common on dynamic construction projects [3,50–55]; (b) global positioning systems (GPS) require expensive investments once high-level accuracy is needed and as such are not suitable yet to be deployed on workforce to track its precise movements [56,57]; (c) GPS and passive/active radio-frequency identification (RFID) system [58,59] have been applied extensively for material tracking at acceptable error rates that practitioners tolerate; (d) laser scanning or photo- or videogrammetric approaches for project progress, as-built status, and earned-value tracking [60–66] provide valuable as-built geometry potential for safety analysis [67]; and (e) multiple sensor/ data fusion-based approaches for tracking equipment operator visibility or workers location and their physiological status data [68–70]. All technologies require eventually intensive infrastructure to collect and process data. Although several of these approaches have the potential to provide real-time location data to resources, the proposed framework uses commercially available (although infrastructure intensive and invasive) wireless (RF) and GPS sensors to record the location of resources in real time (at a 1 Hz update rate) and laser scanners to capture the 3D as-built environment (once, at the time of the experiment). Latter is used to analyze the site geometry, for example, slope ratios in excavation [67], and other temporal construction hazards resulting from tools and equipment used, or hazards built into the design. Selected technologies contribute also to the visualization of the resource location and frequency analysis of near misses.

3. Definitions, objectives, and scope The overall framework of the pursued research is shown in Fig. 1. It presents a holistic view of the work areas that are necessary for automated safety performance monitoring and assessment in construction. The emphasis of this article is on the recording and evaluation of proximity hazard events.

Accurate and emerging remote sensing technology (see box 1 in Fig. 1), with a particular emphasis on real-time detection and tracking of construction resources (workers, equipment, and material), can gather the critical spatial–temporal data that are needed for the purpose of this research [48,71]. Fusion of tracking information generated from remote sensing data, combined with information on safety regulations and site geometry, has the potential to advance the understanding of the actual execution of construction tasks. Seeking quantifiable information as the result of the processing (e.g., checking data quality, detecting and identifying location and size of hazard zones) allows numerous practical avenues that are needed for automated safety performance analysis (e.g., counting and mapping too close proximity of workers-on-foot to equipment) (see box 2 in Fig. 1). New knowledge in construction occupational safety and health is then rapidly created if the automation of monitoring and assessment of the safety performance succeeds (see box 3 in Fig. 1). The objectives of this article are explained in more detail, before details to the design, testing, and validation of the method for construction safety measurement and assessment related to proximity or near miss hazards are presented. The first objective of the selected approach is to automatically identify the primary areas of hazards on a construction site. These are defined: • Static hazards: Hazards inherently built into the construction design or plan, including hazards that result from temporary construction conditions and aids (examples: excavation slope ratios, power, or other site logistical issues such as location of temporary material laydown yards, storage of hazardous substances, construction site traffic control, and physical hazards such as openings on a floor) [2]. • Dynamic hazards: Hazards related to the spatial–temporal movement of resources (examples: workers and ground operating equipment in motion, cranes with loads swinging over a work space) [48,71–75].

The second objective is to automatically gather and analyze the spatial–temporal conflicts between workers-on-foot and the identified hazards. Several typical proximity hazards were considered in the design of the developed method. Although this article focuses mostly on the context of the proximity of workers-on-foot to dynamic hazards, typical proximity hazards include the following cases: • Close proximity of workers with static or dynamic equipment and materials (e.g., machinery and as-built objects) • Falls of workers from higher elevation (e.g., openings and leading edges on a floor or roofs or elevated platforms)

Fig. 1. Data collection, processing, and safety application framework.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

• Proximity to pre-defined areas where chemical, flammable, and toxic substances are present (e.g. containers, high-voltage power lines) • Unauthorized intrusion to access-controlled or restricted (work) spaces (e.g., crane swing areas, access to confined and other limited or restricted work spaces without training or certification, and unstable slopes with cave-in potential in excavated pitches or trenches) • Monitoring the use of designed uniform traffic control devices (MUTCD) or installations (e.g., signs, sidewalks, or pedestrian crosswalks)

The third objective is to define and test a proximity hazard indicator (PHI) that can be utilized to measure and geospatially map near misses and potentially correct the safety performance of the involved stakeholders (incl. the behavior of workers or equipment operators). Near miss information is hereby limited to: • Identify hazardous conditions and worker behavior before a construction accident occurs. This includes in this case: too close interaction (e.g., proximity) of workers-on-foot to construction equipment or geo-referenced hazard areas.

The developed framework is tested on a few selected but common construction activities which, as industry safety statistics have repeatedly shown, experience high to very high injury or fatality rates. The scope of this article is limited to the introduction and initial testing of the developed framework, including the capabilities of the proximity hazard indicator (PHI) that has high potential to function as a leading indicator in the prevention of struck-by events in construction. 4. Methodology of the developed framework The developed method for analyzing the spatial–temporal relationship between workers-on-foot and hazards common on construction sites is illustrated in Fig. 2. The block diagram explains the generation of hazard zones that lead to a spatial–temporal analysis of the existing construction resources on site (e.g., personnel, equipment, and material). Tracking and mapping proximity events ultimately yield a proximity hazard indicator (PHI) that measures the safety performance related to too close proximity of workers-on-foot to equipment (aka. near misses or near hits). Static and dynamic hazards are classified into subcategories: hazards built-into design, temporary static hazard zones, permanent hazards, and dynamic hazards. Although most hazards built-into design might be mitigated already at the design stage or the latest at the construction planning stage, they might still appear on site. Common issues on construction sites

61

are floor openings that are not protected. Examples to temporary static hazard zones are unsecured materials or fenced areas. Permanent hazards are caused by as-built structures. They progress with site activity and become obstacles. Dynamic hazards stem from moving vehicles or loads. According to their spatial–temporal status, all subcategories fall into the main categories of static hazards and dynamic hazards. As data to all hazards are collected, processing in three major steps finally leads to the proximity hazard indicator (PHI): a) A user manually (for now) reviews construction design and plans and generates hazard zones by applying occupational safety and health rules, regulations, and best practices. As a result, static and dynamic hazards zones are pre-identified. If they cannot be mitigated using traditional safety plans or methods, they need to be georeferenced on the site layout plan. Marking includes the type of the construction resource, work task, and schedule. Eventually safety envelopes around the hazards are created to protect the resources. Examples are safe distances to leading edges or restricting a safe work space around heavy construction equipment. This task is already required in the daily job hazard analysis (JHA) procedures on construction sites. Though [2,76,77] have shown, construction hazard detection and mitigation utilizing safety rule checking in building information models (BIM) works already (semi-) automatically to prevent hazards built into the design, they do not resolve any hazards that appear again on site. b) Automatically gather continuous real-time location data to dynamic resources and record and analyze the events when multiple resources intersect a hazard zone (a spatial–temporal breach of the safety envelope of two independent resources). Eventually all potential hazards related to the dynamics of resources are represented by zones featured by the locations and kinetic information of the resources. c) Compute the proximity hazard indicator (PHI) value that evaluates the proximity hazard quantitatively and visualizes results on a site layout plan. Note that the results can be quickly be used again in daily JHA procedures.

The specific technologies and techniques implemented for tracking the construction resources' spatial–temporal data and gathering of the geometries of major objects on construction site have been introduced in the authors' previous research [12,48,78]. These articles focus on tests of 3D as-built modeling from range point clouds, including modeling excavated pits, building concrete and formwork structures, material site storage and assembly facilities, and temporary resources such as

Fig. 2. Block diagram of identifying proximity hazards and measuring proximity hazard indicator (PHI).

62

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

power supply and distribution. They further concentrate on error analysis for continuous location tracking of workers-on-foot and heavy construction equipment. The field tests in these articles demonstrate the general applicability of the technologies for proximity event tracking, detection, identification, and data visualization. 4.1. Definition and generation of hazard zones and safety envelopes A hazard zone is represented as a polygon that is generated based on its geo-referenced location and geometric (size) information. The method that is used to generate a hazard zone varies according to the type of the hazard. A hazard is classified either static or dynamic, as previously explained. A static hazard is either pre-defined according to the construction environment which geometry is known (e.g., accesscontrolled space that only authorized personnel is allowed to enter), or monitored through remote location tracking and sensing technology (e.g., ultra wideband (UWB) or global positioning system (GPS)) [12,48, 78]. The location of a dynamic hazard is gathered utilizing these or other real-time location tracking and sensing technologies. Additional hazards, caused by blind spaces, are introduced as another potential cause for accidents. All of these are explained next in more detail. 4.2. Hazard zones built into design Types of hazards can be assigned to known hazardous construction site conditions before the construction operation starts. Hazard zones built-into design characterize a type of static hazards that are inherently associated to static site geometry and structural components. Examples include but are not limited to one of the following cases: • • • •

openings and leading edges on roofs or elevated platforms, high-voltage power lines, unstable slopes in excavated pits or trenches, and confined and other limited or restricted access or work spaces.

Although a construction schedule is dynamic, above identified hazards all have at minimum temporary, if not fixed locations and geometries (defined as zones). Their as-built geometry is either already known or can be surveyed in regular time intervals using terrestrial [50,62–64,78] or airborne sensing approaches [79]. Detecting hazard locations and boundaries (e.g., planar surfaces or bounding boxes) in the proposed approach is either performed manually or automatically from sensor-based approaches [46–48,78,80] or rule-checking approaches [2].

use or replenished when empty), they can be tagged and their location tracked using sensing technology (e.g., active RFID or GPS tags). The developed approach uses primarily UWB or GPS sensing to flag or find a hazard's location in real time and at high accuracy. Outlier data to erroneous location data were removed through Robust Kalman Filtering like it has been introduced in [57,78,88]. The filtered data were processed to form a polygon that represents the boundary for the hazard.

4.4. Safety envelopes After gathering the specific location and geometry to each hazard, a pre-defined safety envelope is added similar to a more theoretical laboratory-based approach that [49] introduced. The Occupational Safety and Health Administration (OSHA) gives examples in regards to the size of designing a safety envelope. According to OSHA's standard in subpart M 1026.502 [81], “a piece of mechanical equipment cannot be used within 6 feet from the leading edge of a roof,” a 6-feet wide safety envelope along the roof edge must be enforced. The proposed approach uses the OSHA standard but allows modification if a user desires to change the value. The safety envelope is applied once the boundary of the hazard is modeled using existing geometry information. For example, the OSHA standard requires a 5-foot-clear distance of workers to an individual portable flammable liquid tank when the tank's capacity exceeds 1100 gallons [81]. Should information to the safety envelope not been specified in already published safety regulations, a user can specify an appropriate factor based upon the specific situation. Fig. 3 illustrates the modeling of a safety envelope around a static hazard using tags of positioning technology. A series of (dashed black) circles indicate the safe distance (radius) to the hazard polygon (solid blue lines). The circles are centered at the polygon nodes (blue solid dots) as well as points along the edges. A new polygon which creates the safety envelope (red solid lines) around the hazard is formed by connecting the external tangential point of each of the circles. The hazard zone is the area between the safety envelope and the physical hazard object. Eventually a protected hazard has the safety envelope displayed as a bounding box to make the approach economical and less confusing for workers that work nearby or install the physical safety equipment.

4.3. Temporary static hazard zones Another type of hazard is the existence of temporally placed objects that have hazards associated to its use or requirements for secure storage. Such dangers can be formalized as temporary static hazard zones. Because of the dynamic nature of ongoing construction operations, these zones might not be pre-allocated during the project design or planning phases. They may occur rather randomly on site. Examples include but are not limited to the following cases that are present at specific job site locations on an as-need basis: • flammable liquids, such as petrol, alcohol, and welding gas, • chemical and toxic substances, such as acid and alkali solvents, • fire extinguisher and other aids to protect safety and health of workers, and • high-voltage power-generating units, electric wires, and cutting tools. Similar to the hazard zones built-into design, the generation of a static hazard zone requires the knowledge of the hazard location as well as a safety envelope (e.g., diameter around an industrial compressed gas cylinder). Since the temporary static hazard type is of temporal existence in a specific work space (supplied when needed, removed after

Fig. 3. Construction of a tagged static hazard zone.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

4.5. Dynamic hazard zones Construction equipment can rotate on the spot, for example, cranes, excavators, or skid steer loaders. Once in motion, it can also shift into different directions rather quickly. Dynamic hazard zones are characterized by conditions that change over time. Examples are location, shape, scale, and orientation. Examples include but are not limited to one of the following cases: • a worker is walking across a road without using an available crosswalk while a piece of construction equipment or vehicle is nearby; • a worker is performing work tasks around (e.g., behind) a piece of equipment or vehicle while it is in reverse; or • a crane with load swings over a (crew of) worker(s) on the ground or over worker(s) in the blind space of crane operator.

The dynamic hazards (zones) related to a particular piece of equipment can be characterized by four parameters: major function, size, location, and velocity of (parts of) the equipment. The major function of the equipment defines whether it is a piece that operates mainly on the ground and/or also performs lifting tasks. The size of the equipment influences the geometry of the hazard zone. The location of a dynamic hazard is determined by the position where the equipment is located. The velocity of the equipment or parts of the equipment determines the orientation and shape of the generated hazard zone. In contrast to static hazard zones, a dynamic hazard zone does not have a fixed location and velocity. Multiple positioning and orientation sensors (e.g., UWB or GPS tags, or inertial motion units (IMU)) mounted on equipment are able to derive (averaged) location and speed vectors of the equipment. Previous research, i.e. [48,57,88], reported on supplemental information about how many positioning sensors need to be deployed on equipment to provide position and orientation measurements. Four key parameters (function, scale, location, and velocity of the considered equipment) that are determined are used to form a dynamic hazard zone. Fig. 4 shows a piece of ground-moving equipment. The illustration explains how a dynamic hazard zone is generated for a piece of equipment with basic geometry and few or no rotating or

63

translating parts (e.g., skid steer loader, dump truck, pick-up truck, fork lift) [90]. The equipment and its orientation are tracked by several UWB tags, and each tag is mounted on the various parts of the equipment. The position of the equipment is represented by its center point (O in Fig. 4), which is derived by computing the geometric average of the tracking data collected by all tags (assuming the tags are equally located on the equipment). Besides the location, several input parameters are required to generate a hazard zone around dynamic equipment. These parameters include the width (D) and the length (L) of the base of the equipment (assuming the equipment has no swinging, sliding parts or rotating parts); a radius (r) that creates a safe distance (stopping distance including a safety buffer); reaction and braking time (Δt) that the equipment needs to avoid hitting an object; and steering angle (s) when the equipment moves. Knowing these parameters, a polygonal representation of the hazard zone around the equipment is generated through the following procedures: 1. Expand the length and width of the equipment to form a warning zone. A warning zone indicates a clearance area with restricted entrance when the equipment is static. This zone is defined in order to avoid potential injury such as worker being hit by the unexpected movement or rotation of the equipment or parts. 2. Extend the warning zone by a distance of 12 v
t on the equipment moving direction to form the box A–E–F–J. The velocity v is computed through the tracking data, and t is braking time, a given parameter it takes for a full stop of the equipment. Zone A–E–F–J represents the area that can be covered by the equipment during the braking time Δt if the equipment is moving straightforward at the speed of v (assuming flat surface). The braking distance is 12 v
Δt when a linear deceleration model is utilized. 3. Rotate the box A–E–F–J with the angle s both clockwise and counterclockwise about the fixed center at O to form two boxes M–C–D–I and B–G–H–K, respectively. The steering angle is determined by two consecutive vehicle position measurements. Considering the equipment operator may steer the vehicle while braking in order to avoid upcoming objects, the moving direction of the equipment may vary. Boxes M–C–D–I and B–G–H–K indicates the area that can be covered by the equipment if the

Fig. 4. Simplified geometry of the dynamic hazard zone a moving ground vehicle creates (in plain view).

64

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

operator steers on both left and right direction from the very beginning till the equipment stops. 4. Connect nodes A–B–C–D–E–F–G–H–I–J–K–M to form a polygon, which results in the dynamic hazard zone around this piece of equipment. The dynamic hazard zone is generated based on the kinematics and geometric status of the equipment. It is also a prediction of the area that can be covered by the equipment during the braking time. The dynamic hazard zone is stored using the same geo-referenced coordinate system that is used for static hazard zones. Several intermediate parameters are computed using the following equations at the time t when the position of the equipment (in this application: the center of the geometric boundary around the equipment on a 2D projection) is p ¼ ðx; yÞT and the velocity of the equipment is v ¼ ðvx ; vy ÞT : sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi     L vt 2 W 2 R1 ¼ þ rþ þrþ ; where v ¼ jvj 2 2 2 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi     L 2 W 2 rþ þ rþ R2 ¼ 2 2 θ ¼ sin−1

    2r þ L W þ 2r ; θ ′ ¼ tan−1 2R1 L þ 2r

−1

α ¼ tan

  vy ; β ¼ σ þ θ; γ ¼ σ þ θ0 vx

ð1Þ

ð2Þ

ð3Þ

ð4Þ

In a fixed known Cartesian system XOY, the coordinate of each node is computed using the following equations: Denote a vector Z ¼ Z ðφÞ ¼ ðcos φ; sin φÞT

ð5Þ

  F ¼ F x ; F y ¼ p þ R1
Z ðα−θÞ

ð6Þ

  G ¼ Gx ; Gy ¼ p þ R1
Z ðα−σ þ θÞ

ð7Þ

  H ¼ Hx ; Hy ¼ p þ R1
Z ðα−βÞ

ð8Þ

  I ¼ Ix ; Iy ¼ p þ R2
Z ðα−π þ γ Þ

ð9Þ

    J ¼ J x ; J y ¼ p þ R2
Z α−π þ θ0

ð10Þ

    K ¼ K x ; K y ¼ p þ R2
Z α−π−σ þ θ0

ð11Þ

Notice that the dynamic hazard zone is symmetric along the central axis, therefore the coordinates of the rest of the nodes can be computed. Two approximations were made when forming a convex hazard zone. First, the arcs on both front and rear sides of the equipment are replaced by the chords; second, straight edges C–B and H–I are used on both left and right side. Fig. 4 and Eqs. (1)–(11) detail the generation of a dynamic hazard zone for a piece of equipment (e.g., trucks, loaders, dozers) that operates at the same elevation level as workers-on-foot. In contrast, other types of equipment exist that perform articulated motions where equipment parts extend or rotate (e.g., buckets of excavators, hooks of mobile, or tower cranes). These motions can be simplified into translating and revolving. A mobile crane, for example, conducts forward and backward translations, while the revolving part is the superstructure including the boom, the hook, and the load. Equipment with revolving components has two dynamic hazard zones: one centered around the translating component and the other centered on the revolving part. If the substructure or base of the equipment is immobile while the superstructure is operating, the substructure's hazard zone becomes static. 4.6. Permanent hazard zones generated by equipment blind spaces Construction sites generally consist of numerous objects of various sizes and shapes which limit the field of view (FOV) of an equipment operator. Blind spaces (aka. blind spots) are another potential type of a permanent hazard zone that contributes to many accidents. They must be considered when modeling hazard zones. The hazard zone originating from blind spaces in construction typically originates from two scenarios: (1) ground-moving equipment passing by an object that blocks the operator's FOV and (2) equipment with revolving components that limit the operator's FOV when loading objects behind obstacles [46,82]. Fig. 5 illustrates a case for the formation of blind space hazard zones as a piece of ground-level equipment traverses by such an obstacle. It is assumed that the obstacle itself is not hazardous and subsequently does not need a safety envelope. In a second simplified case shown in Fig. 6, an operator turns the upper structure of a piece of rotating equipment. For the example of a mobile crane with a fixed operator cabin close to the rotating center of the upper rotating structure, the blind space remains almost static. Although blind spaces caused by the cabin structure are not considered further in the developed approach, two static hazard zones exist when the load is not moving: around the base of the crane and below/around

Fig. 5. Dynamic hazard zone of a ground-moving equipment with blind spaces generated through an obstacle.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

65

proximity hazard indicator (PHI). The PHI can be expressed as the following: Proximity Hazrd Indicator ðPHIÞ X α  No: of Proximity Events ðEntries by other ResourcesÞ in Hazardous Zone i i i ¼ Total Observing Time ½min

ð12Þ

Fig. 6. Hazard zones of a revolving equipment with blind spaces.

the hook if elevated or positioned on the ground. Dynamic hazard zones from blind spaces are then caused, when the load starts swinging. A dynamic hazard zone below/around the hook is generated equal to the one for moving equipment on the ground. When the crane hook swings inside the blind space caused by a tall obstacle, additional space is added to the hazard zone. The solid red line indicates the dynamic and blind space hazard zones. Assuming that the obstacle does not pose a danger by itself, no safety envelope around its boundary is required. 4.7. Spatial–temporal data gathering and analysis The worker-on-foot and equipment location are tracked in realtime. The hazard zones on construction sites are defined manually. A point-in-polygon algorithm of spatial–temporal analysis was developed to examine whether any resource intrudes a hazard zone at any given moment t, and predicts the close proximity of two resources for a short period at the time t + Δt. A worker's safety status at a current time (t) is determined by the worker's intrusion status at both current moment (t) and next state (t + Δt). This worker is considered to be safe only if (s)he is outside any hazard zone at both current and predicted moment. This analysis can be simplified as a point (worker's location) in a polygon (hazard zone) problem. Alternatively, a buffer (aka. safety envelope) around the worker can keep a safe distance between the worker and the hazard. Numerous algorithms have been developed to deal with the point-in-polygon problem. These algorithms are generally classified into two groups: ray casting algorithm (crossing number algorithm) and winding number algorithm [83]. In this paper, the crossing number method is utilized which counts the number of times a line starting from worker's position crosses the edges of the boundary of a (safety) polygon. A point is outside when this “crossing number” is even; otherwise, when it is odd, the point is inside. The same procedure is repeated on the current and predicted positions of a worker. 4.8. Proximity hazard indicator (PHI) The point-in-polygon algorithm for the spatial–temporal analysis classifies worker's activities and performances into safe and unsafe. A worker's safety performance is further measured by the developed

where i is the index of a hazard zone and αi is the safety factor of each hazard zone. Similar to recent research, a multi-attribute decision making tool can be implemented to formalize the specific safety factors that relate to equipment activity [84]. In our case, αi ≡ 1 in the test bed environment explained of the first experiment (see Section 5.1). The PHI represents how often the observed target is exposed to various defined hazards within a fixed observation period. The PHI can be calculated for an individual, a crew of workers, or multiple pieces of equipment. A typical construction operation may last for hours or even days. Conventional safety inspection methods rely on random and discrete observations, which may result in biased assessment of any safety performance (e.g., workers-on-foot, equipment operators). The PHI is capable to continuously monitor the safety performance of such resources; however, information becomes more meaningful when segmenting the safety inspection of the entire operation into a series of smaller observing periods. A PHI is computed for each observation period (e.g., 5 minutes), which forms a series of PHIs distributed over time. Statistical analysis can be therefore conducted to address the average safety performance for the specific operation, the moment when most of the unsafe behaviors occur, and the individual or piece of equipment who/which frequently violates the safety rules or enters into hazard situations. Since the tags of the location tracking technology also provide the unique identification values to the human subjects involved, linking such information could potentially refer to the name of the worker. How such data, once processed to valuable information, can be used to improve construction site safety is discussed in a later section. 5. Experiment and results The developed proximity hazard detection and assessment method has been tested in three experiments. The first experiment was conducted in a larger but controlled test bed environment where the participants, including workers-on-foot and vehicles, performed various realistic safe and potentially unsafe tasks after a scripted predesigned scenarios (note: all scenarios were safe to the individuals and performed by professionals). The second and third experiment tested the developed data collection and processing methods in a live and uncontrolled construction site (again conducted by professionals). The results and discussions to each experiment are presented accordingly. Although the gathered data were post-processed, research [82, 89] has already shown that valuable real-time feedback to warn construction operators and/or workers can happen instantly. 5.1. Experiment in a test bed with simulated events The first experiment was conducted in an outdoor environment to simulate a material handling scenario (see Fig. 7). This experiment intended to test the performance of the developed method when detecting various types of unsafe proximity cases. The experiment occupied a 35-by-35-m flat ground area without obstacles. Six UWB receivers were set up on the ground plane and a video camera was mounted at elevation to monitor the test bed. Five voluntary subjects were recruited. UWB tags were mounted on the subjects' helmets, while one participant also wore a robotic total station prism to collect ground truth of the positioning data (for measuring the error). 163,007 location data points were recorded within the 6 minutes period

66

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

Fig. 8. Proximity events between a worker-on-foot, a static hazard, and a moving vehicle.

test bed and gathered trajectory data to each subject are color-coded in Fig. 7. A photo was added to increase the visual understanding. 35 unsafe proximity cases were detected. The locations are plotted in the Fig. 7. Two of these cases are explained in more detail:

Fig. 7. Data and proximity events in the test bed with simulated working scenarios (while the colors match the resources' trajectories in the graph above, the photo illustrates the test bed, the static and dynamic resources, and their velocity vectors at the beginning of the experiment).

the experiment lasted. The error analysis according to the method introduced in [48] resulted in a 0.34 m average tracking error and a 0.16 m standard deviation. Two of the subjects were instructed to move several boxes from and to fixed locations in the test bed (start to end; no. 1 and no. 2, respectively). Their route was pre-determined. Two additional subjects simulating equipment were instructed to approach the first two subjects (workers-on-foot) to create near miss situations. Additional UWB tags were used to geo-reference two static objects (hazards no. 1 and no. 2), e.g., red cones and a dolly, to simulate static hazards. The equipment had permission to enter the hazard envelope of the static hazards, while workers-on-foot were not given the permission. Additional location tracking data were recorded by the fifth subject, for example, a series of location data the subject recorded simulated a crane hook's movement swinging into the test bed while the FOV of the crane operator was occluded by a (virtual) wall (that did not exist in reality). The simulated crane hook was positioned at height with an attached load. Obliging to safety rules and best practices, no worker should be present at any time below a crane load. The experimental

(1) One subject (worker-on-foot no. 1), shown in Fig. 8, walked through a static hazard area (small red dots inside purple polygon) on all of the return paths and was twice too close to a dynamic hazard (big red rings inside red polygon) that traversed close to the unloading area. Note that equipment orientation as well as timestamp were recorded. Precise reconstruction of the latter event, including direction of both the worker and the vehicle, is possible. (2) Another subject (worker-on-foot no. 2), shown in Fig. 9, was exposed to a live, swinging crane load. The red triangle with blue fill represents the position of the crane load at the time of the proximity event. The red polygon represents the dynamic hazard zone. The subject was also present in the blind space of the crane operator when the event happened.

Fig. 9. Proximity events between a worker-on-foot to an elevated crane load and within the blind area of a crane operator.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

67

Table 1 Summary of the results related to the simulated working scenarios.

Worker no. 1

Worker no. 2

Proximity events [no.] Duration [mm:ss] Closest distance [m] Duration [mm:ss] Speed [m/s] Proximity events [no.] Duration [mm:ss] Closest distance [m] Duration [mm:ss] Speed [m/s]

Static hazard no. 1

Static hazard no. 2

Equipment no. 1

Equipment no. 2

Crane hook

PHI

0 n/a n/a n/a n/a 0 n/a n/a n/a n/a

10 00:32 1.90 05:17 1.44 1 00:04 2.10 05:41 1.43

2 00:03 2.57 04:25 4.49 0 n/a n/a n/a n/a

11 00:40 12.59 01:54 2.47 7 01:07 3.33 04:49 3.27

2 00:02 2.41 02:36 2.90 2 00:04 1.78 04:01 3.01

4.59

Details to the results of all detected unsafe proximity cases are summarized in Table 1. The first subject had exposed himself to many more proximity events than subject two. Both subjects were too close to static hazard no. 2, equipment no. 2, and the simulated crane hook. Given that the calculated PHI value of subject no. 1 (4.59) is much higher than of subject no. 2 (1.83), consequences might be drawn to pull the individual or both from the operation and provide them with additional safety education and training. Besides such event information, further information to each proximity event can be calculated. This includes, but is not limited to: (a) the duration of the proximity event (the time the subject or equipment spent within the pre-defined hazard zone), (b) the closest distance between a subject or equipment to a static or dynamic hazard, (c) the timestamp the closest distance was recorded, and (d) the speed at the time of the closest distance.

In the event of a static hazard, the closest proximity distance was computed as the distance of the worker to the centroid of the hazard area (assuming the hazard is located at the center, otherwise it would need to be calculated to the closest edge). The absolute speed information was used. In the event of the proximity of a worker-on-foot to a dynamic hazard, the closest proximity distance and the speed represent the relative displacement and movement between the worker-on-foot and the centroid of the hazard area. Extracted data to a time period of 5 minutes and 27 seconds (αi ≡ 1) illustrates the richness of the information. The second subject (workeron-foot no. 2) had 7 unsafe proximity events to a swinging crane hook. As listed in Table 2, a steady but rather short distance was maintained between the subject and the crane hook. Although the height of the crane hook was considered when computing the closest distance, the projection of the center of the crane hook on the ground could be used as well. Furthermore, the entry and exit time indicate the duration

1.83

of the proximity event. The positions of worker and crane hook as well as their relative velocity provide further detail. The results show that it is possible to detect three types of unsafe proximity events: worker-onfoot traverses through static hazard zone, worker-on-foot in too close proximity to a moving vehicle (dynamic hazard), and worker-on-foot is within the blind area of a crane operator. 5.2. Experiment and results in a larger and controlled environment A second, controlled experiment was designed to validate the error and efficiency of the developed proximity detection method that measures the location and time of near miss events. This experiment was conducted on the top level of a parking garage (50 by 110 m in dimensions; see Fig. 10). Multiple UWB tags were deployed on five subjects-on-foot, two dynamic vehicles (sedan cars), and three static hazard zones throughout the test bed environment. Further traffic control was embedded in the test bed, e.g. safe paths for the subjectson-foot to cross the vehicle lanes through crosswalks. The error of the location tracking technology (UWB) was then measured using a robotic total station as introduced by [46,82]. The UWB tracking error had a mean of 0.27 m and standard deviation of 0.31 m. It was concluded that the technology's low error allowed spatial–temporal data analysis for detection proximity events between workers-on-foot and vehicles. Furthermore, three video cameras, observing the entire test bed, recorded the entire test time. Video data were manually analyzed afterwards to accurately validate the detection of proximity hazards. In addition, a professional (human) safety expert was placed in the test bed to function as a site safety inspector. His manual notes were further used to gain insights from manual vs. automated safety performance testing. However, these results are mostly opinion-based. Fig. 10 shows a plan view of the site and the scripted scenarios (walking and driving paths) for the participating subjects in the experiment. A static hazard zone (red polygon) was marked at the north side of the parking garage as a restricted access area (e.g., confined work

Table 2 Details to each of the unsafe proximity interactions between worker no. 2 and equipment no. 2.

Closest distance [m] Duration [mm:ss] Entry time [mm:ss] Exit time [mm:ss] Worker no. 2 position

Crane hook position

Relative speed

X [m] Y [m] Z [m] X [m] Y [m] Z [m] Vx [m/s] Vy [m/s] Vz [m/s] V [m/s]

Proximity event no. 1

Proximity event no. 2

Proximity event no. 3

Proximity event no. 4

Proximity event no. 5

Proximity event no. 6

Proximity event no. 7

13.07 00:09 02:16 02:25 14.55 25.46 0.13 14.69 28.97 12.59 −0.91 −0.41 0.02 2.15

12.91 00:07 02:47 02:54 14.06 25.57 0.16 15.18 28.21 12.59 0.98 −0.19 0.07 3.56

13.03 00:06 03:17 03:23 14.85 25.87 0.24 14.59 29.23 12.59 −0.91 −0.42 −0.05 2.92

12.60 00:09 03:43 03:52 14.20 29.08 0.15 13.94 29.38 12.59 −0.87 −0.50 0.04 3.93

12.72 00:12 04:03 04:15 14.87 29.23 0.08 13.66 30.59 12.59 −0.67 −0.74 −0.04 3.58

12.67 00:14 04:25 04:39 14.91 28.91 0.10 13.70 29.68 12.59 0.72 0.69 0.03 3.13

13.33 00:12 04:49 05:01 15.58 29.77 0.14 12.97 31.09 12.59 −0.55 −0.76 −0.04 3.27

68

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

Fig. 10. Layout of the controlled experiment with various scripted scenarios in a safe test bed.

space area) that was strictly forbidden to be entered. Two cones each with a UWB and unique identifier were positioned on the south side of the parking deck. They represented static hazards of temporal nature (e.g., gas tanks). One crosswalk each was located on both north and south sides of the test bed. The experiment included two (cars) vehicles. Their designated travel paths are indicated by dashed lines. Both vehicle drivers followed pre-instructed/designed travel paths on clockwise and counterclockwise directions. On the north side, both vehicles were instructed to accelerate to approximately 25 km/h (exceeding the typical speed limit of construction sites which is about 5 km/h). The planned trajectories of the five subjects-on-foot are shown in solid lines. They were instructed to perform the following five scripted scenarios during the experiment. Each scenario lasted about 20 minutes. The experiment has been repeated five times such that all subjects rotate in each scenario at least once: • Scenario no. 1 (S1): This subject always walked in an area without overlapping with any pedestrian or vehicle traffic. The subject overall performed only activates on safe paths. • Scenario no. 2 (S2): This subject moved parallel to the traffic lane by keeping a safe distance to the vehicle lanes. This subject, however, frequently walked through a static hazard zone on the north side of the experiment site. • Scenario no. 3 (S3): This subject crossed the traffic lanes without using crosswalk on a regular and frequent basis. • Scenario no. 4 (S4): This subject crossed the traffic lanes using the crosswalk on the north side. This subject also walked on the traffic lane on the east side. • Scenario no. 5 (S5): The subject crossed the traffic lanes with and without using the crosswalk on the south side, and randomly approached the moving vehicle from arbitrary directions. This subject also entered the two static hazard zones frequently.

An additional participant in the study functioned as a safety inspector. He was trained to manually observe the safety performance of all resources in the entire test bed. The safety performance measurement method used was based on a behavior-based safety (BBS) observation technique. The inspector was unaware of the scripted tasks of the subjects and vehicles. Since all the other subjects rotated among the working scenarios, it exposed the inspector to a very realistic environment that might occur on a busy construction site. The observer was instructed to keep a data record of any unsafe behavior he observed

at any of the subjects and vehicles in each of the five scenarios. The BBS technique only recorded the number of participants that were exposed to different hazards without reporting the repetitions of the same participant involved in the same hazard. A trained but independent volunteer participated to post-examine the video clips from the three video cameras. This individual was unaware of the goal of the study and was only trained to analyze the video for counting the proximity events. Although the camera position may not have been ideal to validate every proximity event, the volunteer felt that he was able to clearly distinguish safe from unsafe actions in the test bed in all cases. The volunteer was able to replay the video sequence of a potential proximity incident, while the inspector in the test bed had only one chance to decide whether the incident was safe or unsafe. The reports from the safety inspector and video interpreter were compared to the results gained from the automated proximity detection method for validation. The trajectories of all subjects and vehicles and the locations of the proximity incidents detected by the method for each scenario are plotted in Fig. 11. Such visualization of data alone becomes very useful for practitioners because they can, for the first time, plot where and when near misses occur. Intuitively speaking, subject no. 5 in the scenario relating to the observation period 60–80 minutes had several near misses to static and dynamic hazards. As a minimum result, this subject performed differently than most other subjects performing the same scenario. This may lead to a conclusion whether the subject was skilled enough to execute the task. Several more interesting research questions could be asked and answered with the availability of such data. Consequences could be drawn, for example, installing traffic signals or warning devices, or pulling the subject from performing the task to increase the safety performance. Similar actions might be taken for subjects nos. 3 and 4 in the observation period 40–60 minutes. More objective quantitative data analysis needs to be done. Table 3 summarizes the number of the detected proximity cases that were found by the inspector though BBS, by the video interpreter though the manual analysis of the video clips, and the automated analysis using the developed automated proximity detection method. According to Table 3, the results show that the developed method always detected a greater number of proximity cases than the analysis performed through carefully watching the video clips. The manual video clips provided the ground truth data for detecting unsafe proximity cases (near misses). More details comparing the results achieved by these two approaches are detailed in Table 4.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

69

Fig. 11. Trajectories and detected location of proximity cases in the controlled experiment.

Table 3 Proximity cases detected by the research through automated detection and tracking, video clips analysis, and manual on-site behavior-based safety inspection, and the (A: proposed method, C: video clips, BBS: behavior-based safety). No.

Scenario 1

Scenario 2

Scenario 3

Scenario 4

Scenario 5

Subtotal

Subtotal

A C BBS A C BBS A C BBS A C BBS A C BBS A C BBS A C BBS

Static hazardous zones

Dynamic hazardous zones

Static 1

Static 2

Static 3

Dynamic 1

Dynamic 2

16 16 2 15 12 0 8 6 2 12 12 0 6 5 0 57 51 4 188 174 12

22 21 2 4 4 0 10 8 2 12 10 0 6 5 0 54 48 4

22 22 0 18 18 2 9 9 0 11 10 0 17 16 2 77 75 4

5 5 1 6 6 1 10 10 2 8 7 1 7 6 0 36 34 5 72 65 6

6 6 1 8 7 0 7 6 0 10 7 0 5 5 0 36 31 1

Total

71 70 6 51 47 3 44 39 6 53 46 1 41 37 2 260 239 18 260 239 18

Two types of errors were found after evaluating the videos: missdetection and over-estimation. A miss-detection is a proximity incident that was recognized in the video but was not detected by the method. This type of error was caused by insufficient quality of the tracking data, especially when a subject moved outside the line-of-sight of the UWB sensing infrastructure. In this experiment, four miss-detections were recorded, all related to the dynamic proximity incidents. None occurred in the static incidents. Another type of error was over-estimation, which means that the developed method detected a proximity incident that was considered safe in the manual analysis of the volunteer. Further exploration suggests that the over-estimation should not always be considered as wrong. In some cases, the method was more consistent than the human judgment. For example, it is pretty difficult for a human observer to judge the distance of a subject to a hazard if both are very close to each other, but far away from the observer. This may not even account

Table 4 Validation of the results using video analysis. Total

Manual video analysis

Computational method detection

Safe Unsafe

Safe

Unsafe

N/A 4 (miss-detections)

25 (improve: 21, over-estimate: 4) 235

Precision = 98.3%; recall = 90.3%.

70

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

for any fatigue or distraction a human observer may have during a work day. Since the method relies on automatically generated data that calculate the distance between hazards based on pre-defined parameters, its results are more consistent. The evaluation of dynamic cases can be interpreted similarly. A manual observation using the video may not provide accurate vehicle velocities, especially when the vehicle was close or has been speeding. The method gave more consistent results. In this experiment, 21 overestimations of the static cases were measured due to this reason. This type of over-estimation, however, is much of an improvement to the current manual observation. The other four over-estimations occurred when the vehicle was steering slightly within the lanes at high speed. This is a current limitation of the method. The method generated a dynamic hazard zone by predicting a tangent line along the current moving direction. This hazard zone became less reliable when the vehicle was performing a sharp directional change, since the changes to the velocity vector mainly occur in the lateral direction. The participants traversed among the five experimental scenarios until each participant had been involved in every scenario. The Proximity Hazard Index (PHI) of each participant was calculated for every 2minute interval using Eq. (12), and the results are shown in Fig. 12. It can be noticed from scenarios 4 and 5 that all participants had significantly higher PHI values than in any other scenarios. High PHI values indicate that work tasks performed in these two scenarios exposed the

subjects to more proximity incidents than other in the other tasks. Such data if collected and processed in real-time can be used to take almost real-time decisions and advise corrective actions to improve work environment or conditions. Ultimately, this experiment proves that workers may not be blamed for unsafe behavior, but that often the working or site conditions lead to unsafe behavior. A worker that has to cross a vehicle lane to perform a work task requires a pedestrian crosswalk to safely cross the lane. This was not provided in the experiment. If it is not provided, as shown in the scenarios, near miss events happen. The gathered results further support novel design for safety and post safety performance analysis, including practices with accurate time and location of near misses. 5.3. Experiment and results from a construction site The developed method was finally tested on a real construction site in a third experiment. The details to the setting have already been described in the authors' previous work and are not repeated again [48]. The experiment involved a mobile crane operator (lifting material into a pit) and eleven workers (tying rebar and erecting formwork). Several sensing technologies were deployed in this experiment. A time-of-flight laser scanner gathered very accurate as-built conditions of the observed site's geometry. The collected point cloud data were

Fig. 12. Distribution of the Proximity Hazard Index of all the participants.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

71

Fig. 13. Detecting hazardous conditions and unsafe proximity cases in a construction pit from range point clouds.

processed through a method depicted in the authors' previous work [78]. The illustrations in Fig. 13 show the hazardous conditions identified in the work environment of the workers. Location of unfinished building structures, material lay down areas, and potential tripping hazards such as from temporary power lines (120 V) can be seen (note: the power lines at the construction site were elevated to avoid workers tripping). Other hazards such as a high slope ratio on one of the excavated sites in the pit were not considered since all construction workers had professional safety training and were allowed to enter this classified “confined work space” area. Some of the erected formwork and rebar installation caused blind spaces to

Fig. 14. Distribution of the crew's PHI value computed by the developed method.

the crane operator. Moreover, the same UWB system as in the previous experiment was deployed to collect the spatial–temporal data of the crew and the mobile crane. The average error of the location tracking data in this experiment was 0.34 m using the same validation method as explained before. The developed proximity detection method was utilized to analyze the collected UWB data and the results are plotted in Fig. 13. The workers were in close proximity (less than 6 feet) to the elevated temporary power lines 156 times and were exposed to the crane jib/hook two times within the 56-minute observation period. Since all temporary power lines were elevated (a good safety practice), they posed no harm to the crew. However, should the power cords have lied on the ground – like it is typical for many construction sites – it would have created many potential trip hazards. The PHI of the entire crew was based on 2-minute observation intervals. The result to PHI is illustrated in Fig. 14. 73% of all unsafe proximity incidents occurred within the last 20 minutes of the total length of the experiment. A detailed trajectory analysis to the UWB tags implies that the high frequency of proximity incidents toward the end of the observation period was caused by three material lifts. A mobile crane performed these into the pit area. The first started in the 37th minute and each lasted for about 2–6 minutes. During each hoisting task, the crew on the ground (workers-onfoot) yielded the movement of the crane job and load. Although the workers during these lifting periods kept – as instructed through safety training – a safe and clear distance to the potential lift hazard(s), the confined space they worked in required them to get closer to the elevated temporary power lines and next to the earth embankment with high slope ratio. Although all actions were safe, the PHI values increased due to the proximity to potential risks/hazards embedded in the site conditions. It can be concluded from this experiment that site conditions often play a key role in safety performance!

72

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73

6. Conclusions and outlook Construction safety performance in the past decade has stagnated when only marginal improvements have occurred. This research shows how reporting can be made a positive experience by describing how proximity events leading to near miss data can be effectively collected, analyzed, and used for safety engineering and management. Although not all proximity cases can be counted as near misses (because the term near miss has yet to be defined), in-depth understanding of construction activities through (semi-) automated monitoring and analysis is feasible. By recording and analyzing proximity cases, an additional metric for worker safety performance becomes available and in particular, how data related to near misses can be used to improve safety processes in construction. The presented technical contributions show that advanced real-time location sensing and topographic survey technologies have made it possible to quickly and accurately capture live spatial–temporal data of static and dynamic construction resources in multiple simulated and as-built construction environments. As data to such technologies become available, its analysis through novel computational data mining and processing leads to information that can be used to identify and resolve safety issues. There is high potential that the developed approached will result in a transformation of the current procedures of recordable incident reporting. Information will become much quicker and closer available to actions that happen on real construction sites. In the future, it needs to be the goal to avoid hazards in the first place as researchers [2] have already very effectively shown. Considering in-depth understanding of hazardous proximity incidents between workers-on-foot and equipment that often are inherently built into the design and planning of construction site operations are a logical next step to pursue in research and commercial development. The developed automated computational algorithms are based on the analysis of construction project site geometry, spatial, temporal, and kinematic characteristics of various static or dynamic construction resources. A model has been tested in three different environments and has been validated by comparing to manual observations and video records. The results demonstrate that the developed automation of previous manual efforts can accurately, consistently, and reliably detect and measure a worker-on-foot's safety performance under proximity hazards. The developed approach has potential to assist in measuring at least one key issue in construction site safety: Determining the number of proximity events that lead to struck-by incidents. Using proximity hazard indicator (PHI) as a leading safety indicator, safety performance of each individual (incl. the ratio of the number of proximity events of workers-on-foot type, operator, type of equipment, and site layout conditions) can be taken into evaluation. Once such ratings are available, it is feasible for a construction site safety manager to identify frequently occurring proximity hazards before incidents happen. Actions that prevent incidents can be taken ahead of time. Although the availability of this new PHI so far relies on a classical point-in-polygon algorithm – used to find if a worker is within a hazardous zone – the presented research has demonstrated that the resulting information is capable to generate additional knowledge (i.e., providing near miss data) that indicates safety performance over time. However, it does not measure whether an injury or fatality has occurred or not occurred. We termed information as the results from processed raw data. Safety knowledge is – among other types – further reasoning and application, for example, which driver at what speed caused how many and what type of near misses, and what actions have to follow? Worker-, crew-, and task-specific safety education and training might also occur in the future based on tracked records and PHI performance data. The developed approach yields high potential overcoming some of the drawbacks that exist in today's safety orientation and job

hazard analysis (JHA) [77,87]. In the future, they can become much more engaging by providing site-relevant safety performance information of accurate historic data values [85–88]. This may lead to effectively engage the workforce in the process of traditional hazard awareness programs. Advanced safety knowledge can further be used to improve the effectiveness and efficiency of safety education, training, and review of safety incident reporting programs. Further and more detailed studies are necessary, in particular how to implement technology in the field and receive worker acceptance and buy-in. Technical advancements are necessary in the field of standards, for example, how can erroneous data and uncertainty of the developed technology in the harsh, rapidly changing work environment be eliminated? Moreover, the developed automation utilizes several simplified external parameters, such as equipment breaking time and possible steering angle. Since these parameters are currently defined empirically (based on industry expert opinions and specific machine data), any inaccurate parameter setting may result in unreliable measurement and feedback. Improvement occurs also if these parameters are well defined through further study of construction site layout and traffic analysis. Future data may additionally take automatically recorded machine data into account, such as telemetric systems become common on newer machines. If successful, the developed approach may also be extended to other construction safety and health applications such as worker health and equipment productivity analysis.

References [1] J. Gambatese, S. Rajendran, M. Behm, Green design and construction: understanding the effects on construction worker safety and health”, Prof. Saf. – Journal of the American Society of Safety Engineers 52 (5) (2007) 28–35. [2] S. Zhang, J. Teizer, J.K. Lee, C. Eastman, M. Venugopal, Building Information Modeling (BIM) and safety: automatic safety checking of construction models and schedules, Autom. Constr. Elsevier 29 (2013) 183–195. [3] J. Teizer, Status Quo and Open Challenges in Vision-Based Sensing and Tracking of Temporary Resources on Infrastructure Construction Sites, Advanced Engineering Informatics, Special Issue Infrastructure Computer Vision, Elsevier 29 (2) (2015) 225–238. [4] Bureau of Labor Statistics, National census of fatal occupational injuries in 2010, http:// stats.bls.gov/news.release/pdf/cfoi.pdf2010 (Accessed, July 2, 2012). [5] Bureau of Labor Statistics, Census of fatal occupational injuries charts, 1992–2010, http:// www.bls.gov/iif/oshwc/cfoi/cfch0009.pdf2010 (Accessed, July 2, 2012). [6] Division of Occupational Safety and Health, “Health hazards in construction”, Washington State Department of Labor and IndustriesMay 2011 http://wisha-training.lni.wa.gov/training/presentations/HealthHaz.pps2011 (Accessed, August 23, 2012). [7] Bureau of Labor Statistics, National census of fatal occupational injuries in 2008, http:// www.bls.gov/news.release/archives/cfoi _08202009.pdf2009 (Accessed, July 30, 2012). [8] J.W. Hinze, J. Teizer, Visibility-related fatalities related to construction equipment, J. Saf. Sci., Elsevier 49 (5) (2011) 709–718. [9] J. Teizer, B.A. Allread, C.E. Fullerton, J. Hinze, Autonomous pro-active real-time construction worker and equipment operator proximity safety alert system, Autom. Constr., Elsevier 19 (5) (2010) 630–640. [10] J. Teizer, Magnetic field worker proximity detection and alert technology for safe heavy construction equipment operation, Proceedings of the 32nd International Symposium on Automation and Robotics in Construction, Oulu, Finland, 2015. [11] J. Teizer, Safety 360: surround view and alerts to protect workers from hazardous equipment, Proceedings of the 32nd International Symposium on Automation and Robotics in Construction, Oulu, Finland, 2015. [12] T. Cheng, J. Teizer, Modeling tower crane operator visibility to minimize the risk of limited situational awareness, J. Comput. Civ. Eng. 28 (3) (2014) 04014004. [13] Occupational Safety and Health Administration, Hazardous and toxic substances, http://www.osha.gov/SLTC/hazardoustoxicsubstances/index.html2012 (Accessed, August 23, 2012). [14] Occupational Safety and Health Administration (OSHA), “Fire protection and prevention – flammable liquids”, Safety and Health Regulations for Construction. Part-1926.152, http:// www.osha.gov/pls/oshaweb/owadisp.show_document?p_id=10673&p_table=STANDARDS2012 (Accessed, September 18, 2012). [15] L. Levine, “Worker safety in construction industry: the crane and derrick standard”, Congressional Research Services Report for CongressReleased on November 21, 2008 http://www.fas.org/sgp/crs/misc/RL34658.pdf2008 (Accessed, August 23, 2012). [16] National Safety Council, Injury facts2008 edition. Itasca, IL, ISBN 978-0-87912-278-2 https://www.usw12775.org/uploads/InjuryFacts08Ed.pdf2008 (Accessed, July 30, 2012). [17] R. Hefner, “Construction vehicle and equipment blind area diagrams”, National Institute for Occupational Safety and Health, Report No. 200-2002-00563, http://origin.cdc.gov/ niosh/topics/highwayworkzones/BAD/pdfs/catreport2.pdf2004 (Accessed May 11, 2008). [18] D.E. Fosbroke, Studies on heavy equipment blind spots and internal traffic control, Roadway Work Zone Safety & Health Conference, Baltimore, MD2004. [19] J.S. Ray, J. Teizer, Coarse head pose estimation of construction equipment operators to formulate dynamic blind spots, Adv. Eng. Inform., Elsevier 26 (1) (2012) 117–130. [20] R. Bostelman, J. Teizer, S.J. Ray, M. Agronind, D. Albanesee, Improved methods for evaluation of visibility for industrial vehicles towards improved safety standards, Saf. Sci., Elsevier (2014) 257–270.

J. Teizer, T. Cheng / Automation in Construction 60 (2015) 58–73 [21] S.J. Ray, J. Teizer, Computing 3D blind spots of construction equipment: implementation and evaluation of an automated measurement and visualization method utilizing range point cloud data, Autom. Constr., Elsevier 36 (2013) 95–107. [22] T.M. Ruff, Evaluation of devices to prevent construction equipment backing incidents, National Institute for Occupational Safety and Health, November 20042004. [23] T.M. Ruff, T.P. Holden, Preventing collisions involving surface mining equipment: a GPSbased approach, J. Saf. Res. 34 (2003) 175–181. [24] E. Marks, J. Teizer, Method for testing proximity detection and alert technology for safe construction equipment operation, Construction Management and Economics, Taylor & Francis, Special Issue on Occupational Health and Safety in the Construction Industry31(6)2013 636–646. [25] T.M. Toole, Construction site safety roles, J. Constr. Eng. Manag. 128 (3) (2002) 203–211. [26] J. Hinze, R. Godfrey, An evaluation of safety performance measures for construction projects, J. Constr. Res. 4 (1) (2003) 5–15. [27] A. Carbonari, A. Giretti, B. Naticchia, A proactive system for real-time safety management in construction sites, Autom. Constr. 20 (6) (2011) 686–698. [28] A. Giretti, A. Carbonari, B. Naticchia, M. DeGrassi, Design and first development of an automated real‐time safety management system for construction sites, J. Civ. Eng. Manag. 15 (4) (2009) 325–336. [29] K. Tantisevi, B. Akinci, Simulation-based identification of possible locations for mobile cranes on construction sites, J. Comput. Civ. Eng. 22 (1) (2008) 21–30. [30] American Bureau of Shipping, Guidance notes on safety culture and leading indicators of safetyJanuary 2012 http://www.eagle.org/eagleExternalPortalWEB/ShowProperty/BEA% 20 Repository/Rules&Guides/Current/188_Safety/Guide2012 (Accessed, July 6, 2012). [31] A. Laufer, W.B. Ledbetter, Assessment of safety performance measures at construction safety, J. Constr. Eng. 112 (4) (1986) 530–543. [32] American National Standards Institute, Method of recording and measuring injury experienceZ-16.1 American National Standards Institute, New York, 1973. [33] J. Hinze, C. Harrison, Safety programs in large construction firms, J. Constr. Div. 107 (3) (1981) 455–467. [34] United Kingdom Health and Safety Executive, A guide to measuring health and safety performance, http://www.hse.gov.uk/opsunit/perf meas.pdf2001 (Accessed, July 2, 2012). [35] J. Hinze, Leading indicator of construction safety performance, Presented on ConocoPhillips Downstream Construction Network – Face to Face Meeting, Los Angeles, CA, September 20, 2006, 2006. [36] Cambridge Center for Behavior Studies, “Introduction to behavioral safety”, http://www. behavior.org/resource.php?id=3302012 (Accessed, July 6, 2012). [37] J.S. Chhokar, J.A. Wallin, Improving safety through applied behavior analysis, J. Saf. Res., Elsevier 15 (4) (1984) 141–151. [38] D.P. Fang, X.Y. Xie, H.L. Huang, Factor analysis-based studies on construction workplace safety management in China, Int. J. Proj. Manag., Elseveir 22 (1) (2004) 43–49. [39] H. Lingard, S. Rowlinson, Behavior-based safety management in Hong Kong's construction industry, J. Saf. Res., Elsevier 28 (4) (1997) 243–256. [40] R.M. Choudhry, D.P. Fang, Why operatives engage in unsafe work behavior: investigation factors on construction site, Saf. Sci., Elsevier 46 (4) (2008) 566–584. [41] M.R. Hallowell, J.W. Hinze, K.C. Baud, A. Wehle, Proactive construction safety control: measuring, monitoring, and responding to safety leading indicators, J. Constr. Eng. Manag. 139 (10) (2013) 04013010. [42] D. Zohar, G. Luria, The use of supervisory practices as leverage to improve safety behavior: a cross-level intervention model, J. Saf. Res., Elsevier 34 (5) (2003) 567–577. [43] J. Agnew, A. Daniels, Development high-impact leading indicator for safety. The Performance Management Magazine, Accelerating Business Performance, http://aubreydaniels.com/ pmezine/developing-high-impact-leading-indicators-safety2011 (Accessed, July 6, 2012). [44] A.L. Glendon, S.G. Clarke, E.F. McKenna, Human safety and risk management, CRC Press, Taylor and Francis Group, 2006, ISBN 0-8493-3090-4. [45] S. Gadd, A.M. Collins, Safety culture: a review of the literature, http://www.hse.gov.uk/research/hsl_pdf/2002/hsl02-25.pdf2002 (Accessed, July 6, 2012). [46] J. Teizer, C. Kim, C.T. Haas, K.A. Liapi, C.H. Caldas, A framework for real-time 3D modeling of infrastructure, Transp. Res. Rec.: Journal of the Transportation Research Board 1913 (2005) 177–186 (Washington D.C.). [47] J. Teizer, C.H. Caldas, C.T. Haas, Real-time three-dimensional occupancy grid modeling for the detection and tracking of construction resources, J. Constr. Eng. Manag. 133 (11) (2007) 880–888. [48] T. Cheng, M. Venugopal, J. Teizer, P.A. Vela, Performance evaluation of ultra wideband technology for construction resource location tracking in harsh environments, Autom. Constr., Elsevier 20 (8) (2011) 1173–1184. [49] X. Luo, W.J. O'Brien, F. Leite, J.A. Goulet, Exploring approaches to improve the performance of autonomous monitoring with imperfect data in location aware wireless sensor networks, Adv. Eng. Inform. 28 (4) (2014) 287–296. [50] M. Golparvar-Fard, A. Heydarian, J.C. Niebles, Vision-based action recognition of earthmoving equipment using spatio-temproal features and support vector machine classifiers, Adv. Eng. Inform. 27 (4) (2013) 652–663. [51] J. Yang, T. Cheng, J. Teizer, P.A. Vela, Z.K. Shi, A performance evaluation of vision and radio frequency tracking methods for interacting workforce, Adv. Eng. Inform. 25 (2011) 736–747. [52] J. Yang, P.A. Vela, J. Teizer, Z.K. Shi, Vision-based crane tracking for understanding construction activity, J. Comput. Civ. Eng. 28 (1) (2014) 103–112. [53] S. Chi, C.H. Caldas, Automated object identification using optical video cameras on construction sites, Comput. Aided Civ. Infrastruct. Eng. 26 (5) (2008) 368–380. [54] S. Chi, C.H. Caldas, Image-based safety assessment: automated spatial safety risk identification of earthmoving and surface mining activities, J. Constr. Eng. Manag. 138 (3) (2011) 341–351. [55] M.W. Park, C. Koch, I. Brilakis, Three-dimensional tracking of construction resources using an on-site camera system, J. Comput. Civ. Eng. 26 (4) (2011) 541–549. [56] A.A. Oloufa, M. Ikeda, H. Oda, GPS based wireless collision detection of construction equipment, Proceedings of 19th ISARC, National Institute of Standards and Technology, Gaithersburg, Maryland, September 23-25 2002, pp. 461–466. [57] N. Pradhananga, J. Teizer, Automatic spatio-temporal analysis of construction site equipment operations using GPS data, Autom. Constr., Elsevier 29 (2013) 107–122. [58] J. Song, C.T. Haas, C.H. Caldas, E. Ergen, B. Akinci, Automating the task of tracking the delivery and receipt of fabricated pipe spools in industrial projects, Autom. Constr. 15 (2) (2006) 166–177.

73

[59] D. Grau, C.H. Caldas, C.T. Haas, P.G. Goodrum, J. Gong, Assessing the impact of materials tracking technologies on construction craft productivity, Autom. Constr. 18 (7) (2009) 903–911. [60] J.C. Du, H.C. Teng, 3D laser scanning and GPS technology for landslide-earthwork volume estimation, Autom. Constr., Elsevier 16 (2007) 657–663. [61] S. El-Omari, O. Moselhi, Integrating 3D laser scanning and photogrammetry for progress measurement of construction work, Autom. Constr., Elsevier 18 (1) (2008) 1–9. [62] P. Tang, D. Huber, B. Akinci, R. Lipman, Automatic reconstruction of as-built building information models from laser-scanned point clouds: a review of related techniques, Autom. Constr. 19 (7) (2010) 829–843. [63] F. Bosché, Automated recognition of 3D CAD model objects in laser scans and calculation of as-built dimensions for dimensional compliance control in construction, Adv. Eng. Inform. 24 (1) (2010) 107–118. [64] J. Gong, C.H. Caldas, Data processing for real-time construction site spatial modeling, Autom. Constr. 17 (5) (2008) 526–535. [65] M. Golparvar-Fard, J. Bohn, J. Teizer, S. Savarese, F. Peña-Mora, Evaluation of image-based modeling and laser scanning accuracy for emerging automated performance monitoring techniques, Autom. Constr. 20 (8) (2011) 1143–1155. [66] G. Zhang, P.A. Vela, I. Brilakis, P. Karasev, A sparsity-inducing optimization based algorithm for planar patches extraction from noisy point-cloud data, J. Comput. Aided Civ. Infrastruct. Eng., Wiley 30 (2) (2015) 85–102. [67] J. Wang, S. Zhang, J. Teizer, Geotechnical and safety protective equipment planning using range point cloud data and rule checking in building information modeling, Autom. Constr., Elsevier 49 (2014) 250–261. [68] T. Cheng, J. Teizer, G. Migliaccio, U. Gatti, Automated task-level activity analysis through fusion of real-time location sensors and worker's thoracic posture data, Autom. Constr., Elsevier 29 (1) (2013) 24–39. [69] T. Cheng, G.C. Migliaccio, J. Teizer, U.C. Gatti, Data fusion of real-time location sensing (RTLS) and physiological status monitoring (PSM) for ergonomics analysis of construction workers, J. Comput. Civ. Eng. 27 (3) (2013) 320–335. [70] J. Teizer, B.S. Allread, U. Mantripragada, Automating the blind spot measurement of construction equipment, Autom. Constr., Elsevier 19 (4) (2010) 491–501. [71] S.N. Razavi, C.T. Haas, Multisensor data fusion for on-site materials tracking in construction, Autom. Constr. 19 (8) (2010) 1037–1046. [72] L. Kuo-Liang, C.T. Haas, Multiple heavy lifts optimization, J. Constr. Eng. Manag. 122 (4) (1996) 354–362. [73] A. Shapira, M. Goldenberg, “Soft” considerations in equipment selection for building construction projects, J. Constr. Eng. Manag. 133 (10) (2007) 749–760. [74] Z. Lei, H. Taghaddos, J. Olearczyk, M. Al-Hussein, U. Hermann, Automated method for checking crane paths for heavy lifts in industrial projects, J. Constr. Eng. Manag. 139 (10) (2013) 04013011. [75] Z. Lei, S. Han, A. Bouferguène, H. Taghaddos, U. Hermann, M. Al-Hussein, Algorithm for mobile crane walking path planning in congested industrial plants, J. Constr. Eng. Manag. 141 (2) (2014) 05014016. [76] S. Zhang, K. Sulankivi, M. Kiviniemi, I. Romo, C.M. Eastman, J. Teizer, BIM-based fall hazard identification and prevention in construction planning, Saf. Sci., Elsevier 72 (2014) 31–45. [77] S. Zhang, F. Boukamp, J. Teizer, Ontology-based semantic modeling of construction safety knowledge: towards automated safety planning for job hazard analysis (JHA), Autom. Constr., Elsevier 52 (2015) 29–41. [78] T. Cheng, J. Teizer, Real-time resource location data collection and visualization technology for construction safety and activity monitoring applications, Autom. Constr., Elsevier 23 (2013) 3–15. [79] S. Siebert, J. Teizer, Mobile 3D mapping for surveying earthwork projects using an unmanned aerial vehicle (UAV) system, Autom. Constr., Elsevier 41 (2014) 1–14. [80] H. Rashidi, I. Brilakis, P.A. Vela, Monocular videogrammetry for generating absolute-scale 3D point cloud data of civil and infrastructure scenes, Proceeding of the 2013 ASCE International Workshop on Computing in Civil Engineering 2013, p. 5810588. [81] Occupational Safety and Health Administration (OSHA), “Fall protection. Construction Safety and Health Outreach Program”, U.S. Department of Labor, OSHA office of Training and Education, May 1996, http://www.osha.gov/doc/outreachtraining/htmlfiles/ subpartm.html1996 (Accessed, September 17, 2012). [82] C. Kim, C.T. Haas, K.A. Liapi, C.H. Caldas, Human-assisted obstacle avoidance system using 3D workspace modeling for construction equipment operation, J. Comput. Civ. Eng. 20 (3) (2006) 177–186. [83] P. Heckbert, Graphics Games IV, Academic Press Professional, Boston, 1994, ISBN 0-12336156-7 (1994). [84] A. Shapira, M. Simcha, Measurement and risk scales of crane-related safety factors on construction sites, J. Constr. Eng. Manag. 135 (10) (2009) 979–989. [85] J. Teizer, O. Golovina, D. Wang, N. Pradhananga, Automated collection, identification, localization, and analysis of worker-related proximity hazard events in heavy construction equipment operation, Proceedings of the 32nd International Symposium on Automation and Robotics in Construction, Oulu, Finland, 2015. [86] S. Zhang, J. Teizer, N. Pradhananga, Global positioning system data to model and visualize workspace density in construction safety planning, Proceedings of the 32nd International Symposium on Automation and Robotics in Construction, Oulu, Finland, 2015. [87] S. Zhang, J. Teizer, N. Pradhananga, C.M. Eastman, Workforce location tracking to model, visualize and analyze workspace requirements in building information models for construction safety planning, Autom. Constr. (2015) http://dx.doi.org/10.1016/j.autcon. 2015.09.009 (in press). [88] N. Pradhananga, J. Teizer, Cell-based Construction Site Simulation Model for Earthmoving Operations using Real-time Equipment Location Data, Visualization in Engineering, Springer Verlag 3 (12) (2015) http://dx.doi.org/10.1186/s40327-015-0025-3. [89] J. Teizer, Wearable, wireless identification sensing platform: self-monitoring alert and reporting technology for hazard avoidance and training (SmartHat), J. Inf. Technol. Constr.(ITcon) 20 (2015) 295–312. [90] J.P. Zhang, Z.Z. Hu, BIM and 4D based integrated solution of analysis and management for conflicts and structural safety problems during construction: principles and methodologies, Autom. Constr. 20 (2011) 155–166.