You are on page 1of 24

1

The Less-than-perfect Driver: A Model of Collision-inclusive Carfollowing Behavior


Wuping Xin, PhD Candidate Research and Development Division KLD Associates, Inc 47 Mall Drive Suite 8 Commack, NY 11725 Tel: 631-543-6500-227(ext) Fax: 631-543-4330 Email: wupingxin@kldassociates.com John Hourdos, PhD Minnesota Traffic Observatory University of Minnesota 500 Pillsbury Dr. SE Minneapolis, MN 55455 Tel: 612-626-5492 Fax: 612-626-7750 E-mail: hourdos@umn.edu

Panos Michalopoulos, Professor Department of Civil Engineering University of Minnesota


500 Pillsbury Dr. SE Minneapolis, MN 55455 Tel: 612-625-1509 Fax: 612-626-7750 E-mail: micha001@umn.edu Gary Davis, Professor

Department of Civil Engineering University of Minnesota


500 Pillsbury Dr. SE Minneapolis, MN 55455 Tel: 612-625-2598 Fax: 612-626-7750 E-mail: dtrips@umn.edu

Submitted for Presentation and Publication Transportation Research Board 2008 Annual Meeting January 2008 Washington, D.C. TOTAL WORDS: 6250 + 250 5 = 7500

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

ABSTRACT
In this paper a new behavioral car-following model is presented. Unlike traditional car-following models which preclude vehicle collisions, the proposed model aims to emulate less-thenperfect everyday driving, while striving to capture both safe and unsafe driver behavior. Most importantly, a realistic perception-response process is incorporated with the model based on developments from visual perception studies. In particular, driver inattention is characterized by a driver-specific variable called scanning interval. This variable when coupled with the drivers visual perception-response process, actually results in variable reaction times that are not only dependent on each drivers individual characteristics, but also on instantaneous traffic conditions such as speed and density. This enables closer emulation of real life human driving and its interactions with surrounding vehicles. Stated otherwise, both inter- and intra-driver variations in reaction time are captured in a plausible and coherent manner, while in earlier studies, reaction time is either presumed fixed, or of limited variability. Furthermore, parameters of this model have a direct physical and behavioral meaning; this implies that vehicle collisions, if any, can be analyzed with respect to behavioral patterns rather than simply being treated as numerical artifacts. In all, 54 detailed and accurate vehicle trajectories extracted from 10 real life crashes were employed to test the models capability of replicating freeway rear-end collisions. In addition high resolution crash-free trajectory data were employed to validate the model against normal driving behavior. Test results indicate that the proposed model is able to replicate both normal as well as unsafe driving behavior that could lead to vehicle collisions. Finally, the feasibility of integrating the proposed model with existing micro-simulators is briefly discussed. The outcome of this work can eventually facilitate studying crash mechanisms at a high definition microscopic level as well as enabling safety-related system design improvements and evaluation through micro-simulation software.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

INTRODUCTION
In recent years, considerable research has been conducted for the development, implementation, and evaluation of innovative ITS technologies aiming to improve traffic operations and safety. As part of this process, micro-simulation has become an increasingly indispensable tool for system design and evaluation. However, while being successful in many applications, microsimulation has been insufficient for safety-related studies, i.e., its applications have long been limited to collision-free scenarios. This is because existing micro-simulators employ carfollowing and lane-changing models that are limited to the emulation of normal driver behavior. Collisions are deliberately precluded as they are hard to model. To be sure, realistic collisioninclusive car-following models pertinent to both safe and unsafe driving behavior are still lacking. In recent years, limited research efforts have focused on enhancing existing micro-simulation models by developing conflict statistics and safety surrogates (Archer and Kosonen 2000; Hunuenin et al. 2005; Kosonen 1999; Kosonen and Ree 2000; Ledoux and Archer 1999; Mehmood et al. 2001; Torday et al. 2003). However, due to the basic assumption that followers will always maintain a safe distance, it has been difficult to test and validate the correlation between extracted conflict statistics/surrogate measures with actual crash potential. This inevitably limits the reliability of safety-performance conclusions obtained from crash-free micro-simulation models. In this paper, a new behavioral collision-inclusive car-following model is presented. This model should eventually help to enable micro-simulation as an effective and reliable tool to analyze crash-prone designs and aid in safety evaluations. Unlike traditional car-following models that deliberately prohibit vehicle collisions, the proposed model aims to replicate both safe and unsafe driver behavior. This is accomplished by employing a realistic perceptionresponse process based on developments in visual perception studies. Furthermore driver inattention is characterized by a driver-specific variable called scanning interval. This variable, when coupled with the perception-response process actually results in variable reaction times not only dependent on individual driver characteristics, but also instantaneous traffic conditions such as speeds and density. This allows closer emulation of real life human driving and its interactions with the surrounding vehicles. Model parameters have a direct physical and behavioral meaning which allows vehicle collisions to be introduced as a result of behavioral patterns rather than re-determined simple numerical artifacts. 54 vehicle trajectories extracted from 10 real life crashes and near-crashes were employed to test the models capability of replicating freeway rear-end collisions. High resolution crash-free trajectory data were also used to validate the model against collision-free driving behavior. Test results indicate that the proposed model is able to replicate both normal as well as unsafe driving behavior that could lead to vehicle collisions. Major contributions of this research include: a dataset of over 700 hours of video recordings and high-resolution collision-inclusive vehicle trajectories collected from a high crash rate freeway section of I-94 WB in Minneapolis, Minnesota. The trajectories include individual vehicle position, speed, and acceleration at 0.1 sec resolution ( 1.5 ft accuracy) for a total of 54 vehicles that were involved in 6 freeway crashes and 4 near-crashes. To the best of the authors knowledge, this is one of the first and most elaborate data collection efforts devoted in the collection of vehicle trajectories from real-life crashes to aid car-following modeling. In

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

4 addition, an enhanced perception-response process is integrated into the model based on research results from visual perception studies. This essentially results in variable reaction times dependent not only on driver characteristics but also instantaneous traffic conditions. State otherwise, both inter- and intra-driver variations in reaction time are captured in a plausible and logical manner, while in earlier studies, reaction time is either presumed constant, or of limited variability. Finally, the proposed model was tested using both non-crash and crash data. This should assist in better understanding of unsafe car-following behavior and provide insights for crash-mechanisms at the microscopic (individual vehicle) level.

BACKGROUND
Car-following models have been the subject of numerous research efforts since the 1950s. The motivation arose from the need to model individual vehicle movements as realistically as possible. In conjunction with lane-changing algorithms, car following modeling forms the cornerstone of microscopic simulation tools and is of particular importance for studying rear-end collisions. Briefly, most of existing car-following models can be represented as: an (t + ) = f (x, v, v, a, a; t ) (1) or an (t + ) = f (x, v, v, a, a; t ) + (2) with x(t ) = xn 1 (t ) ln 1 xn (t ) v = vn 1 (t ) vn (t ) a (t ) = an 1 (t ) an (t ) where an 1 is acceleration of lead vehicle; an is acceleration of following vehicle; a represents relative acceleration between lead and following vehicles; ln 1 is the length of lead vehicle; ln is the length of following vehicle; n 1 represents the index of lead vehicle; n represents the index of following vehicle; is the reaction delay of following vehicle; vn 1 is the instantaneous speed of lead vehicle; vn is the instantaneous speed of following vehicle; v is relative speed between lead and following vehicles; xn 1 is the position of lead vehicles bumper; xn is the position of following vehicles bumper; x is the bumper to rear distance between lead and following vehicles; is noise term. These models can be either deterministic as described by Equation (1), or stochastic as in Equation (2) where a certain noise term is added, making them discontinuous in time. In addition, taking the driver and vehicle as one integrated unit, most current car-following models incorporate reaction time delay into the dynamics. This reaction time delay is defined as the time lag between a stimuli effected by the lead vehicle (e.g., braking) and the response by the subject vehicle. Usually it encompasses perception, recognition, decision time plus neural-muscle response time, sometimes also includes vehicles mechanical response delay. Typically, most car-following models assume constant reaction delay time for all drivers, while implemented in some recent studies is the notion of variable reaction time that varies with different types of drivers, or of limited variability depending on discrete driving regimes or traffic situations (Caliper 2007; Rong et al. 2006; Toledo 2003; TSS 2007; Wang 2005). Beyond the classical car following paradigms described above, there is also a notable line of developments termed

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

5 psycho-physical modeling that cannot be solely represented by Equation (1) or (2). Specifically, this type of models assumes vehicle acceleration is more or less constant until a certain actionpoint is reached (therefore, these models are sometimes referred to as action point models). Todesiev (Todesiev 1963) was among the first using action points to characterize car-following process; while the most successful action-point models include Wiedsmann (Wiedemann and U.Reiter 1992) and Fritzsche model (Fritzsche 1994); variants of these two models are implemented in two widely-used commercial micro-simulators, i.e., VISSIM and PARAMICS , respectively. As pointed out earlier, existing car-following models have been successful at reproducing many observed features of traffic flow and normative driving behavior. However, when vehicle collisions and safety become important, these models cannot capture intricacies of individual driver behavior. Specifically, performance limitations, judgment errors, and variable reaction times that account for both intra- and/or inter-driver variations are not well integrated or simply treated as additive random noise. Further, lapse of attention is yet to be properly addressed in a logical and coherent way, and at best treated as random delayed reaction to an event in front. For instance, in the VISSIM simulator, diver inattention is artificially generated by a predefined probability and added as additional reaction delay (PTV 2007). In this paper, a new model is proposed aiming to replicate both safe and unsafe driver behavior while taking into account the weakness and risks of human driving. To that end, a critical modeling effort is on incorporating a realistic perception-response process based on developments in visual perception studies. As will become clearer later, this actually results in both intra- and inter-driver variations in reaction times, hence allowing closer emulation of real life human driving and its interactions with the surrounding traffic dynamics. Once again it should be stressed that classical car-following models have been developed on a crash-free basis, under the assumption (explicitly or implicitly) that the following vehicle always tries to maintain a safety distance from the leader. Albeit in the past years considerable research efforts have been devoted to investigate the so-called instability of classical carfollowing models (e.g., the GM-family models), and actually identified certain unstable parameters that could result in collision situations (Castillo et al. 1994; Gazis et al. 1959; Gazis et al. 1961; Herman et al. 1959; Wilson 2001; Zhang and Jarrett 1997), still crashes are not their intended modeling objectives while vehicle collisions generated with unstable parameters are more of modeling outliers rather than logical products of behavioral patterns.

MODEL
Figure 1 (a) illustrates the conceptual framework of the proposed model. Essentially, the longitudinal driving tasks are described as a negative-feedback control process between the external world and the driver-vehicle-unit (DUV). As shown in this figure, a DUV is an encapsulation of driver and vehicle as one integrated unit. This means a DUV possesses both driver-specific characteristics such as perception limits, and vehicle-specific features such as mechanical delay. Based on the conceptual framework, a specific model is developed in this study. Before elaborating on the modeling structure, key parameters of this model and their implications are first explained below. All these parameters are specific to individual drivers.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

Key Parameters and Implications


1) Maximum Comfortable Deceleration Rate ( ft / s 2 ) bn This is the maximum deceleration rate an individual driver is willing to use in non-emergency situations. It is bounded by the vehicles physical limits for decelerating. Usually this ranges from 9~15 ft/s^2 (0.20 ~0.35 g). Particularly, the Traffic Engineering Handbook by ITE (ITE 1999a) suggests 10 ft / sec 2 as comfortable deceleration rate. Meanwhile, the American Association of State Highway and Transportation Officials (AASHTO) recommends 11.2 ft / sec 2 as the value for comfortable deceleration rate for most drivers (AASHTO 1990). Regardless, in this research, the maximum comfortable deceleration rate is not determined a priori, but as a result from model calibration subject to the feasible box constraints (i.e., simple inequality constraints). 2) Maximum Comfortable Acceleration Rate ( ft / s 2 ) an This is the maximum acceleration rate a driver is willing to use to achieve desired free-flow speed or to catch up with the lead vehicle. This parameter is bounded by the vehicles physical limits for accelerating. 3) Perceptual Threshold of Visual Expansion Rate (radian/sec) C& This parameter characterizes the mechanism a driver uses to identify relative motion. Specifically, the visual stimulus to detect relative motion has been found in visual perception studies to be related to a continuously changing optic array characterized by the visual angle & subtended by the image of the lead vehicle on the retina, and the expansion rate of this angle (Evans and Rothery 1974; Gibson 1982; Lee 1976; Lee and Reddish 1981; Michaels 1965; & Michaels and Cozan 1963; Sun and Frost 1998). Mathematically, and can be specified as (See Figure 1 (b) ):
W /2 tan( ) = 2 D (t )
sin( ) D (t ) = & V

(3) (4)

where

: Visual angle subtended by the image of the lead vehicle on retina; & : Angular velocity, i.e., the expansion rate of ;
W : Vehicle width; D (t ) : Distance headway; V : Relative speed.
When is very small, Equation (3) and (4) can be approximated as:

W D

(5)

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

7
D (t ) = & V

(6) (7)

& =W

V D2

& In order to perceive relative motion, the angular velocity must be above a certain threshold value C& . This means, in order for a following driver to perceive relative velocity in relation to the preceding vehicle,
W | V | D
2

C& must be satisfied. Different values of this threshold

have been reported in the literature ranging from 1 ~ 40 104 radian / sec (Brackstone et al. 2000; Evans and Rothery 1973; Evans and Rothery 1977; Hoffmann 1968; Michaels and Cozan 1963; Wiedemann and U.Reiter 1970s). It is very important to point out that the perception process characterized by this perceptual threshold actually entails a logical and coherent mechanism for variable reaction time. This implication is best explained through the following example. Assuming a driver (vehicle) has a characteristic perceptual threshold of 8 10 4 radian / sec . Both the subject vehicle and the preceding vehicle are traveling with a constant speed of 90 ft/sec and distance headway 180 ft. The width of the preceding vehicle is 5 ft. At time t1 = 0sec , the preceding vehicle starts braking at -10 ft/s 2 . It is not until t2 = 2.1 sec (when W
V (t ) D (t ) 2

= 8 10 4 ) that the following vehicle

begins to perceive the relative motion. This means the subject vehicles actual reaction time to the braking event of the preceding vehicle is at least 2.1 sec. Note t2 varies as a function of initial instantaneous speeds as well as distance headway. In other words, the actual reaction time of the subject vehicle is dependent both on the instantaneous speeds as well as on local density. Considering in real life traffic the instantaneous speeds and distance headways are actually timevarying, the visual perception process described above essentially provides a logically coherent and plausible mechanism that can better explain variable reaction times observed in reality. Most importantly, this mechanism, with its mathematical simplicity can be easily implemented in carfollowing models without the need to store large amount of historical data. Typically, implementing variable reaction times, even with limited variability, requires storing historical data of vehicles past states thus has significant performance impacts. 4) Weber Ratio CD This parameter characterizes another mechanism a driver uses to identify relative motion when the relative speed is very small. Essentially, Weber Ratio refers to the Just Noticeable Change of Distance (JNCD) when the subject vehicle is approaching the vehicle in front (Levinson 1998). This means the following driver cannot perceive any relative motion until the ratio of accumulated change of distance headway to the original distance headway exceeds the Weber Ratio threshold. Note, Weber Ratio CD and Perceptual Threshold of Visual Expansion Rate (radian/sec) C& described above capture the perceptual limits of human driving, thus reflecting the safety potential of the driver. Otherwise stated, a driver with smaller C& and CD can detect

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

8 relative motion more promptly than a driver with greater C& and CD , i.e., the former driver has an effectively smaller reaction time than the latter. 5) Oscillation Acceleration ( ft / s 2 ) aosc and Oscillation Deceleration Rates ( ft / s 2 ) bosc When a driver can not perceive any relative motion between the subject vehicle and the vehicle in front, he will use oscillation acceleration and deceleration rates to maintain the distance headway in relation to the front vehicle. Essentially these two parameters are necessary to capture the oscillatory spirals that have long be recognized as typical car-following behavior. 6) Desired Following Gap Time (sec) t g This is the following gap time (time headway) a driver wishes to maintain. This parameter characterizes the aggressiveness and safety potential of the driver. For instance, a less-aggressive driver would prefer greater following time headway to allow for sufficient reaction time in case of emergent braking, while an aggressive driver would prefer shorter time headway and choose to follow the lead vehicle closely. 7) Error of Desired Following Gap Time g This parameter refers to the percentage error in relation to the actual time headway when a driver evaluates the latter. 8) Scanning Interval (sec) t scan Scanning interval is the time interval a driver samples his/her front view situation while driving. This reflects the attention level and alertness of the driver. Scanning interval is a driver-specific as well as traffic-situation specific parameter. 9) Mechanical Response Delay (sec) td Mechanical response delay is the time delay involved in the drivers implementation of a control decision, i.e., the time lag between the drivers start of the control action and the time such control action takes effect. Note that a unique feature of the proposed model is the notion of effective reaction time. Although there is no explicit reaction time parameter associated with each driver-vehicle-unit, the perceptual thresholds as well as scanning intervals plus mechanical response delay effectively render a variable reaction time that is dependent on instantaneous speed and local density. In other words, reaction time is an implicit time-dependent variable embedded in the model logic. This provides greater flexibility and modeling realism regarding the drivers perception response process than conventional approach assuming fixed reaction time.

Modeling Details
The proposed model follows the logic depicted in Figure 1. As is shown in this figure, DUV samples traffic information from its surrounding environment through the Information Acquisition Component, compiles sampled information to determine its current state, and compares with its desired state. This desired state is derived from the Individual Driving Strategy

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

9 Component. Then a control decision is determined by the Decision Making Component in terms of acceleration or deceleration to minimize any deviation between its current state and desired state. Finally, control decisions are implemented via the Vehicle Mechanics Component after a mechanical response delay. These four components compose the proposed DUV model. Note that situational factors do not include multilane effects, while the impacts of personal factors and environmental factors on drivers perception-decision process are not explicitly incorporated in the model. Once more experimental data and relevant research results from human factors or psychology studies become available, the model can be expanded to take into account these missing factors.

Information Acquisition Component (IAC) The Information Acquisition Component (IAC) samples surrounding traffic environment every scanning interval. The inputs to this component include subject vehicles instantaneous speed, acceleration, preceding vehicles speed and acceleration, and relative speed and distance headway between the subject vehicle and the preceding vehicle. IAC translates these inputs into three variables: & Visual Expansion Rate Change of Distance Instaneous Time Gap Visual Expansion Rate is computed as in Equation(7). Change of Distance is simply the difference between current distance headway and the distance headway at previous scanning interval. Instaneous Time Gap is derived by dividing current distance headway by subject vehicles instantaneous speed. Driving Strategy Component (DSC) The specific driving strategy is to maintain a desired following gap time t g subject to safety constraints. Due to human drivers inability to keep exact following gap time, it is further assumed that a driver will allow an error of g while maintaining his/her desired following gap
time. This means, if the instantaneous gap time is within [(1 g )t g , (1 + g )t g ] , the driver would simply identify it with t g .

Decision Making Component (DMC) Decision Making Component (DMC) determines an appropriate control maneuver (acceleration or deceleration or cruising) based on the information from Information Acquisition Component and Individual Driving Strategy Component. Specifically, at each scanning interval, DMC does the following Boolean checking: 1) True or false: Visual Angle Expansion Rate exceeds the threshold C&
2) True or false: Change of Distance exceeds the threshold CD 3) True or False: Subject DUV is driving with its gap time outside of [(1 g )t g , (1 + g )t g ] If all of the three checking result false, then the DUV is in a subconscious driving state without motivation for either accelerating or decelerating. This is because in this state, the driver is unable to perceive any relative motion ( all perception thresholds not exceeded) while he/she is

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

10 driving with a comfortable following gap time thus there is no need to accelerate to catch up with preceding vehicle or decelerate to avoid a looming collision. However, small displacements of gas pedal, wind or pavement friction could still effect an actual small acceleration or deceleration for the subject vehicle. Usually this rate is in the range of 0.5~0.8 ft / s 2 (Brackstone et al. 2000; Brackstone et al. 2002; McDonald et al. 1997; Sultan et al. 2004; Wang 2005). In this study, it is termed as Oscillation Acceleration/Deceleration Rate because apart from having the preceding vehicle changes its speed drastically, the subject vehicle would keep employing this small rate until its perceptual threshold is exceeded and then the subject vehicle will reverse the sign of the acceleration rate until the perceptual threshold is exceeded again. When neither of the above checking results true, Gipps-like car following rules are used:
a b vn (t + t g ) = min vn (t + t g ), vn (t + t g )

(8) (9)

a vn (t + t g ) = vn (t ) + 2.5an t g [1

vn (t ) v (t ) ] 0.025 + nmax max Vn Vn


v 2 n 1 (t ) } b
n 1

b vn (t + t g ) = t g bn + t g 2bn2 + bn {2[ xn 1 (t ) xn (t ) s0 )] t g vn (t ) +

(10)

where t g is the drivers desired following gap time;

an is the maximum comfortable acceleration rate for vehicle n; bn is the maximum comfortable braking rate for vehicle n; x n x n 1 , are the position of vehicle n and n-1, respectively;

bn 1 is nth drivers estimation for (n 1)th vehicle maximum comfortable braking rate;
Vnmax is the desired free flow speed of vehicle n; s0 is the distance headway at standstill. Essentially Equations (8)(9)(10) reflect the speed the subject driver wishes to achieve within his/her desired following gap time. In these equations, t g is added as time delay rendering the process a time-delay process. Note this is different from the original Gipps car-following model which uses drivers fixed reaction time as the time delay term. To be sure, t g in Equation (8)(9)(10) cannot be interpreted as drivers reaction time; rather it represents a relaxation time during which the driver wishes to relax his/her current speed to a safe speed in order to catch up with the preceding vehicle while avoiding a potential collision. This actually captures the aggressiveness, or time-varying safety potential of the subject driver, because the driver updates his safe speed based on his/her desired following gap time rather than his/her actual reaction time, while the actual value of the latter is varying with different traffic situations.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

11

Vehicle Mechanics Component (VMC) A suitable acceleration or deceleration rate determined by the Decision Making Component (DMC) will be implemented by Vehicle Mechanics Component after a mechanical response delay td . Variable Reaction Time As mentioned earlier, a unique feature of the proposed model is the actual variable reaction time inherently implemented for each individual driver. When combined with scanning interval t scan and mechanical response delay td , a drivers actual reaction time to an event in front becomes a bounded time-varying variable: td actual reaction time Nt scan (11)
where N = 1, 2, 3, depending on at which scanning interval the perceptual thresholds are exceeded.

TESTING
DATA
Two high resolution (0.1 sec) trajectory datasets are used in testing the proposed model. The first is collision-free vehicle trajectories that were collected in a test track using Real-time Kinetic GPS (RTKTPS) by Hokkaido University, Japan (Ranjitkar 2004). The unique feature of RTKGPS dataset is that the lead vehicles speed profiles were carefully controlled during each run to emulate various real-life car-following situations. The other is crash trajectories collected in this study from a high crash-rate freeway section of I-94WB in Minneapolis, Minnesota. Figure 2 (a) gives an aerial view of this section. The location is a three-lane section of approximately 1900 feet in length. Average daily volume is approximately 80,000 vehicles per direction. Figure 2 (b) provides a summary of the 2002 crash statistics both for this section as well as other high crash-rate locations in the Twin Cities area (Hourdos 2005). As shown in Figure 2 (b), the regional average crash rate in the Twin Cities area is 0.96 million-vehicle-miles (MVM) in 2002, whereas the selected section has the highest 4.81 MVM as compared to the other locations. This is almost four times higher than the regional average level and effectively equivalent to one crash every two days. Specifically, inasmuch as crashes are rare and random events, in order to maximize the number of crashes collected, an integrated video collection system was developed in the Minnesota Traffic Observatory (MTO) and deployed taking advantage of wireless communication technology. This integrated system includes CCTV surveillance cameras, on-site small-factor PCs, and wireless communication infrastructure. CCTV cameras were deployed on the roof of a high-rise near the selected data collection site, monitoring and recording the traffic during May and August 2006 from 8:00am to 8:00pm every weekday. Analog videos from these cameras are digitized and compressed at 10fps by the small-factor PCs that are deployed on-site with the cameras. Digitized recordings are transmitted to a remote supervising station through

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

12 the wireless communication infrastructure. At the supervision station, crash recordings are inspected to identify any crashes or near-crashes that have occurred during the day. Identified crash recordings are processed to extract raw trajectories by using a video-based vehicle tracking program called NG-VIDEO (Kovvali et al. 2007). It should be stressed that several non-trivial subtasks have to be manually accomplished prior to applying the NG-VIDEO program. These subtasks include Camera Calibration, World Coordinate Matching, and Video Orthorectification. Camera Calibration means obtaining the intrinsic parameters (focal length, aspect ratio, radial distortion etc.) of the surveillance cameras. Within the context of this study, this involves using the subject camera to snap a planar pattern shown at various orientations, and estimating the cameras intrinsic parameters from the snapshots. Moreover, World Coordinate Matching is the process of mapping pixel coordinates from the crash recordings to world coordinates in real life. This is accomplished with the aid of the ArcGIS software and high-resolution ( 0.5 foot pixel image resolution) ortho-aerial images. Finally, Video Orthorectification requires registering the images of crash videos to the coordinate system in the aerial image. After orthorectification, terrain features in the video images are removed and vehicles appear to be moving on a plane surface. Once these subtasks are accomplished, the outputs are fed to NG-VIDEO. The latter generates raw trajectories data and stores them in a MySQL database. Out of 700 hours of video recordings collected, the final crash dataset includes trajectories of 54 vehicles involved in 6 crashes and 4 near-crashes that occurred at the site from May to August 2006. These trajectories are at 0.1 second resolution, and each trajectory is about 90 seconds. Figure 2(c) illustrates the extracted trajectories of 8 vehicles that had been involved in one such crash. The crash occurred on the rightmost lane between Portland Ave and the flyover to I-94 Westbound at 16:15:48 pm, July 12, 2006. Note that the numbering of vehicle IDs is not necessarily in ordinal order; while in this figure, Vehicle 02 is the second, Vehicle 07 the third vehicle in the platoon and Vehicle 30 is the last one. Similar plots for other crashes/near-crashes are available but not presented here due to space limitations. Detailed information about the crash data collection and processing methodology can be found in (Xin et al. 2007b)

MODEL CALIBRATION
Parameters of the proposed model (represented by a parameter vector ) were calibrated by minimizing the following objective function:

model x sim (t ), x obs (t ) | =

x
t =1

obs

(t ) x sim (t | )
2 N sim

x
t =1

(12)
2

obs

(t ) + x (t | )
t =i

where

is parameter vector; N is total number of observations; x obs (t ) is the time series of observed trajectory for the subject vehicle; x sim (t | ) is the time series of simulated trajectory data for the subject vehicle;

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

13 The objective is equivalently the sum of Theils U statistics with respect to vehicle position data. Generalized Reduced Gradient (GRG) and Branch and Bound approaches were implemented for searching solutions that minimize the objective function. Also the optimization is started with different starting parameters to increase the possibility of finding global minima (for each individual vehicle trajectory, 5000 different initial starts were employed which was found to give a good balance between computational time and convergence of the solution). The feasible region of parameter space is constrained with the lower-bound and upper-bound of each parameter determined from physical limits. For example, a feasible range for Maximum Comfortable Acceleration would be around 11 ft / s 2 (ITE 1999). Finally, the calibration is conducted for each following vehicle, i.e., the observed lead vehicle trajectory was employed as fixed input to the model, and the parameters of the model are optimized so that the modelgenerated trajectory for the follower is as close as possible to the observed data. It should be noted that often an objective function like Equation (13) is used for calibration. However, pooling of data as in Equation (13) for calibration may add new dimensions of difficulty for global optimization, because the observations are correlated (as they are sequential position outputs of car-following model). In addition, there is also a potential problem of heteroscedasticity, i.e., the error associated with the predicted position will depend on its value. Therefore, Theils U statistics as expressed in Equation (12) is used as the objective function for calibration, where the error is somehow normalized to offset the above mentioned effects.

MODEL VALIDATION
Measures of Effectiveness The purpose of validation is to test the models capability of replicating real-life scenarios with calibrated parameters. The following metrics are used to evaluate the deviation between model generated trajectories and real-life trajectories: 1) Root Mean Square Error (RMSE)

RMSE =
where

1 N

( x y )
i i i =1

(13)

xi is the model predicted value indexed by i; yi is the actual observation indexed by i; N is the number of total observations. 2) Mean Percentage Error (MPE)

MPE =
Validation against Non-Crash Data

1 N

| yi xi | xi i =1

(14)

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

14 The test track RTKGPS car-following data are used to validate the model against normal carfollowing behavior. The validation was accomplished by calibrating each DUV using only part of its real-life trajectory but validating it using the entire trajectory in full length. For example a DUV in the platoon has 15 minute-long trajectory at 0.1 sec resolution. One-minute data out of this entire 15 minutes trajectory is used to calibrate the DUV individually, and then the entire 15 minutes trajectory is used to validate the model in platoon situation.

Validation against Crash Data Validation against crash data was conducted by testing if the model can generate the crash event by collective simulation. This means, each DUV is calibrated individually with its preceding vehicles observed trajectory as the input; then all DUVs involved in a crash (usually 5~6 vehicles with the last two directly crashing each other) are simulated collectively with only the platoon leader using observed trajectory as the predefined input. Other vehicles in the platoon are simply updated sequentially using their respective calibrated parameters. If the crash event can be generated as occurred in reality, then the validity of the model is verified.

RESULTS
Figure 3 (a) illustrates the simulation results of non-crash trajectories. In this figure, seven vehicles indexed from G03 to G10 were simulated using the proposed model. Solid black lines represent ground-truth vehicle trajectories while dotted colored lines stand for trajectories predicted by the model. As shown in this figure, the model can replicate theses trajectories fairly well. Additionally, Table 1(a) summarizes the calibrated DUV parameters. The goodness-of-fit measures between real trajectories and model-predicted trajectories are summarized in Table 1(b). The measures presented in the table suggest the model can replicate the trajectories and the speed variations closely. As mentioned earlier, in this study a total of 54 vehicle trajectories from 10 crashes/nearcrashes were used to test the model. However due to space limitations, only one crash case is presented here. The model is equally successful in replicating other crash cases given wellcalibrated parameters. Interested readers should refer to (Xin et al. 2007a) for more detailed results. Specifically, the presented crash, i.e., Vehicle 30 colliding into Vehicle 16, occurred at 16:15pm, July 12, 2006 on the leftmost lane of the I94WB freeway section. At the time the weather conditions were dry and traffic was slow-and-go (see Figure 2(c) ). Figure 3 (b) depicts the simulated crash trajectories vis--vis real crash trajectories. In this figure, the black solid lines are real-life trajectories while the colored dotted lines represent model predicted trajectories. As shown in this figure, the proposed model successfully replicated the crash occurrence. Additionally Table 2 presents the statistics of calibrated parameters values for all 54 crash trajectories. More detailed discussion about the distribution of the calibrated parameter values for this and other crash cases can be found in (Xin et al. 2007a).

CONCLUDING REMARKS
Micro-simulation has become an increasingly indispensable tool for assisting in system design and evaluation in ITS deployment and safety improvements. However, existing micro-simulation models are deficient when evaluating sophisticated safety-related ITS techniques. This is because

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

15 existing micro-simulation models only describe normative car-following behavior devoid of weaknesses and risks associated with real-life everyday driving. To date, realistic car-following models pertinent to the less-then-perfect nature of driver behavior are lacking. In this paper a new car-following model is presented which is capable of more realistically emulating real-life car-following behavior with its risks and imperfections. The model capitalizes on real-life crash data collected from a US freeway as well as test track car-following data from Hokkaido University in Japan to assist model development, calibration and validation. Test results based on actual crash data indicate that the model is capable of replicating both normal driving behavior and unsafe behavior that could lead to collisions. Prior to closing, it should be noted that due to space limitations a detailed discussion of some important issues related to model calibration/validation, e.g., parameter identifiability and sensitivity, as well as in-depth analysis of the parameters distributions are not presented. Also, the models capability to generate well-known macroscopic traffic phenomena such as shockwaves, the fundamental diagram, is not examined in this paper. Such topics are beyond this papers initial scope and are the subjects of several others currently under way. Furthermore, research efforts to integrate the proposed model with existing commercial simulators such as AIMSUN or VISSIM are in progress. These simulators provide flexible APIs that can be used to facilitate integrating the proposed model. However, since the proposed model is collisioninclusive, once integrated with a full-featured simulator environment, more issues need to be resolved before the simulator can be employed in actual safety evaluations or safety-related applications. One critical issue would be potentially long simulation running time or significantly increased number of replications in order to get meaningful safety indicators. A potential solution to this is the employment of parallel computing techniques. Finally it is hoped that this research will contribute our understanding of car-following behavior while improving microsimulation modeling to facilitate assessing freeway safety concepts at a high definition microscopic level.

ACKNOWLEDGEMENTS
This study utilized the advanced detection and surveillance stations maintained by the Minnesota Traffic Observatory while support for this study was provided by the ITS Institute of the University of Minnesota. The authors wish to thank Professor Takashi Nakatsuji and Dr. Prakish Ranjitkar from Hokkaido University, Japan for providing the RTKGPS data.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

16

REFERENCE AASHTO. (1990). A Policy on Geometric Design of Highways and Streets, American Association of State Highway and Transportation Officials, Washington D.C. Archer, J., and Kosonen, I. (2000). "The potential of micro-simulation modeling in relation to traffic safety assessment." ESS Conference Proceedings Hamburg. Brackstone, M., Sultan, B., and McDonald, M. (2000). "Some findings on the approach process between vehicles on motorways." Transportation Research Record, 1724, 21-28. Brackstone, M., Sultan, B., and McDonald, M. (2002). "Motorway driver behavior: studies on car following " Transportation Research Part F:Traffic Psychology and Behaviour, 5, 329-344. Caliper. (2007). "TransModeler 1.0 User Manual." Castillo, J. M. D., Pintado, P., and Benitez, F. G. (1994). "The reaction time of drivers and the stability of traffic flow." Transportation Research Part B:Methodological, 28(1), 35-60. Evans, L., and Rothery, R. (1973). "Experimental measurements of perceptual thresholds in carfollowing." Highway Research Record, 464, 13-29. Evans, L., and Rothery, R. (1977). "Perceptual threshold in car-following- A comparison of recent measurements with earlier results." Transportation Science, 11(1), 60-72. Evans, L. C., and Rothery, R. W. (1974). "Detection of the sign of relative motion when following a vehicle." Human Factors, 16, 161-173. Fritzsche, H. T. (1994). "A model for traffic simulation." Traffic Engineering and Control, 35(5), 317-321. Gazis, D. C., Herman, R., and Potts, R. B. (1959). "Car-following theory of steady-state traffic flow." Operations Research, 7(4), 499-505. Gazis, D. C., Herman, R., and Rothery, R. W. (1961). "Nonlinear follow-the-leader models of traffic flow." Operations Research, 9(5), 545-567. Gibson, J. J. (1982). "A theoretical analysis of automobile driving." Reasons for Realism, E. Reed and R. Jones, eds., Lawrence Erlbaum Associates, Hillsdale,NJ, 199-136. Herman, R., Montroll, E. W., Potts, R. B., and Rothery, R. W. (1959). "Traffic dynamics: Analysis of stability in car following." Operations Research, 7(1), 86-106. Hoffmann, E. R. (1968). "Detection of vehicle velocity changes in car-following." Proceedins:Australian Road Research Board, 4, 821-837. Hourdos, J. (2005). "Crash prone traffic flow dynamics: Identification and real-time detection," Doctoral Dissertation, University of Minnesota, Minneapolis. Hunuenin, F., Torday, A., and Dumont, A.-G. (2005). "Evaluation of traffic safety using microsimulation." 5th Swiss Transport Research Conference (STRC), Monte Verita/Ascona. ITE. (1999). Traffic Engineering Handbook, Institute of Transportation Engineers, Washington D.C. Kosonen, I. (1999). "HUTSIM-Urban traffic simulation and control model: Principles and applications," PhD Dissertation, Helsinki University of Technology, Helsinki. Kosonen, I., and Ree, S. "The potential for microscopic simulation in traffic safety and conflict studies." Proceedings of Conference on Road Safety on Three Continents, Pretoria, South Africa. Ledoux, C., and Archer, J. (1999). "Assessing the safety impact of intelligent transport systems." Royal Institute of Technology, Stockholm http://www.infra.kth.se/ctr/publikationer/ctr1999_08.pdf

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

17 Lee, D. N. (1976). "A theory of visual control of braking based on information about time-tocollision." Perception, 5, 437-459. Lee, D. N., and Reddish, P. E. (1981). "Plummeting gannets: a paradigm of ecological optics." Nature, 293, 293-294. Levinson, W. H. (1998). "Interactive highway safety design model: Issues related to driver modeling." Transportation Research Record, 1631, 20-27. McDonald, M., Brackstone, M. A., Sultan, B., and Roach, C. (1997). "Close following on the motorway: initial findings of an instrumented vehicle study." the 7th Vision in Vehicles Conference, Marseille, France. Mehmood, A., Saccomanno, F., and Hellinga, B. (2001). "Simulation of road crashes by use of systems dynamics." Transportation Research Record, 1746, 37-46. Michaels, R. M. (1965). "Perceptual factors in car following." Proceedings of the Second International Symposium on the Theory of Traffic Flow: London 1963, J. Almond, ed., The Organization for Economic Co-operation and Development (OECD), Paris. Michaels, R. M., and Cozan, L. W. (1963). "Perceptual and field factors causing lateral displacement " Highway Research Record, 25. PTV. (2007). "VISSM 4.30 User Manual." Planung Transport Verkehr AG. Ranjitkar, P. (2004). "Experimental analysis of car following dynamics based on RTK GPS data." Proceesings of Infrastructure Planning 2004, Japan Society of Civil Engineers (JSCE). Rong, J., Wang, L., and Xu, X. (2006). "A variable response time lag analysis using car following data in urban expressway sections." CD-ROM 06-1611 Transportation Research Board Annual Meeting, Washington D.C. Sultan, B., Brackstone, M., and McDonald, M. (2004). "Evidence for the use of deceleration/acceleration in car following." Transportation Research Record, 1883, 3139. Sun, H., and Frost, B. J. (1998). "Computation of different optical variables of looming objects in pigeon nucleus rotundus neurons." Nature Neuroscience, 1(4), 296-303. Todesiev, E. P. (1963). "The action point model of the driver vehicle system," Doctoral Dissertation, Ohio State University Toledo, T. (2003). "Ingegrated driving behavior modeling," Doctoral Dissertation, Massachusetts Institute of Technology. Torday, A., Baumann, D., and Dumont, A.-G. (2003). "Indicator for microsimulation-based safety evaluation." 3rd Swiss Transport Research Conference, Monte Verita. TSS. (2007). "AIMSUN NG Manual v5.1.7." TSS-Transport Simulation Systems. Wang, J. (2005). "A merging model for motorway traffic," Doctoral Dissertation, The University of Leeds. Wiedemann, R., and U.Reiter. (1970s). "Microscopic traffic simulation:the simulation system MISSION background and actual state." PTV Vision Wiedemann, R., and U.Reiter. (1992). "Microscopic traffic simulation:the simulation system MISSION, background and actual state." Project ICARUS(V1052) Final Report, Brussels,CEC. Wilson, R. E. (2001). "An analysis of Gipps's car-following model of highway traffic." IMA Journal of Applied Mathematics, 66(5), 509-537. Xin, W., Hourdos, J., and Michalopoulos, P. (2007a). "Enhanced Micro-simulation models for accurate safety assessment of traffic management ITS Solutions." Project Report,

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

18 Institute of Intelligent Transportation Systems, Center for Transportation Studies, University of Minnesota, Minneapolis. Xin, W., Hourdos, J., and Michalopoulos, P. (2007b). "A Vehicle Trajectory Collection and Processing Methodology and Its Implementation to Crash Data." paper submitted to Transportation Research Board 2008 Annual Meeting, Washington D.C. Zhang, X., and Jarrett, D. F. (1997). "Stability analysis of the classical car-following model." Transportation Research Part B:Methodological, 31(6), 441-462.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

19
TRB 2008 Annual Meeting CD-ROM Paper revised from original submittal.

Figure 1 (a). Conceptual Framework of the Proposed Car-Following Modeling

20

Figure 1 (b). Visual angle subtended by the image of an object on the retina and its expansion rate.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

21

Figure 2 (a). Data collection site: a high crash-rate freeway section of Interstate-94 WB

Figure 2(b) Ten Highest Crash Sections in 2002 (Source: Mn/DOT Crash Facts 2002)

Figure 2(d). Trajectories of 7 vehicles in a crash happened on 07-12-2006

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

22

Figure 3 (a). Validation results using collision-free data (lead vehicle is G03 not shown in the figure)

Figure 3 (b). Simulated trajectories (dotted line) vs. real trajectories (solid line) for Crash Case CR-01
This collision occurred at 16:25 pm, July 12, 2006. Vehicle 30 collided into Vehicle 16 during stop-and-go traffic.

TRB 2008 Annual Meeting CD-ROM

Paper revised from original submittal.

23
TRB 2008 Annual Meeting CD-ROM Paper revised from original submittal.

Statistics
Mean

Standard Deviation

Standard Error of the Mean

25% Percentile

75% Percentile

InterQuartile Range

Minimum

Maximum

Range

Median

Variance

Coefficient of Variance

Kurtosis

Parameters
Max Comft Accel Max Comft Decel Osc Accel Osc Decel

an bn

9.47927 -15.19791 0.98065 -0.83362 1.28736 0.52333 0.87778 0.16667 0.00421

2.40595 3.75494 0.55385 0.1814 0.48314 0.28516 0.54263 0.0866 5.6044E-4

0.80198 1.25165 0.18462 0.06047 0.16105 0.09505 0.18088 0.02887 1.8681E-4

7.0303 -17.3094 0.59419 -0.8792 1.21169 0.29351 0.6 0.1 0.00361

11.3775 -14.1512 0.95337 -0.7552 1.62736 0.83923 1.2 0.2 0.00466

4.3472 3.1582 0.35918 0.124 0.41567 0.54572 0.6 0.1 0.00104

6.01744 -21.4175 0.5095 -1.1163 0.23443 0.11743 0.1 0.1 0.00343

11.7485 -7.95936 1.96256 -0.5977 1.95112 0.88589 1.9 0.3 0.00485

5.73106 13.45815 1.45306 0.5186 1.71669 0.76846 1.8 0.2 0.00142

11.1362 -15.6733 0.75521 -0.8127 1.26694 0.52513 0.8 0.1 0.00432

5.78859 14.09957 0.30675 0.03291 0.23342 0.08132 0.29444 0.0075 3.1409E-7

0.25384 -0.24707 0.56478 -0.2176 0.3753 0.5449 0.61818 0.51962 0.13305

-2.00512 1.34193 0.23958 -0.38262 2.59624 -1.50472 0.37272 -1.07937 -1.70277

aosc bosc

Desired Fol Gap t g Gap Error

Scanning Interval t scan Mechnical Delay t d Expansion Rate Thres C &

Table 1(a) Statistics of calibrated parameters for non-crash data(RTKGPS data): sample size 9

Table 1(b). Validation Statistics (G03 is platoon leader)

24
TRB 2008 Annual Meeting CD-ROM Paper revised from original submittal.
Statistics
Mean

Standard Deviation

Standard Error of the Mean

25% Percentile

75% Percentile

InterQuartile Range

Minimum

Maximum

Range

Median

Variance

Coefficient of Variance

Kurtosis

Parameters
Max Comft Accel Max Comft Decel Osc Accel Osc Decel

an bn

6.92531 -13.7712 0.87709 -1.0489 1.1504 0.42476 0.52292 0.2625 0.0028

1.95146 4.11654 0.25366 0.36772 0.76785 0.24519 0.46184 0.20999 0.00121

0.28167 0.59417 0.03661 0.05308 0.11083 0.03539 0.06666 0.03031 1.7524E-4

5.2196 -17.4338 0.6983 -1.30114 0.56342 0.23924 0.2 0.1 0.00161

7.62579 -11.2517 1.06468 -0.92893 1.25661 0.60246 0.7 0.3 0.00381

2.40619 6.1821 0.36638 0.37221 0.69319 0.36322 0.5 0.2 0.0022

5.00569 -17.99 0.50042 -1.49995 0.5 0.01397 0.1 0.1 0.00101

11.7824 -5.11 1.4479 0.922 3.49641 0.86882 1.9 0.8 0.00491

6.77671 12.8852 0.94748 2.42195 2.99641 0.85485 1.8 0.7 0.0039

6.14407 -15.07755 0.81979 -1.05869 0.96833 0.40122 0.35 0.2 0.00293

3.80821 16.94589 0.06434 0.13522 0.5896 0.06012 0.21329 0.0441 1.474E-6

0.28179 -0.29892 0.2892 -0.35058 0.66746 0.57724 0.88319 0.79996 0.43365

0.01718 -0.82675 -0.41101 17.13454 1.88519 -0.97086 0.63199 0.26576 -1.26009

aosc bosc

Desired Fol Gap t g Gap Error

Scanning Interval t scan Mechnical Delay t d Expansion Rate Thres C &

Table 2. Statistics of calibrated parameters for crash data( I94 crash trajectories) Sample size: 54 drivers

You might also like