Professional Documents
Culture Documents
A R T I C L E I N F O A B S T R A C T
Keywords: In recent years, identifying road users’ behavior and conflicts at intersections have become an essential data
Intersection safety source for evaluating traffic safety. According to the Federal Highway Administration (FHWA), in 2020, more
SSM than 50% of fatal and injury crashes occurred at or near intersections, necessitating further investigation. This
Video analytic tool
study developed an innovative artificial intelligence (AI)-based video analytic tool to assess intersection safety
Extreme value theory
using surrogate safety measures and non-compliance behavior. To extract the trajectory data, a real-time AI
detection model - YOLO-v5 with a tracking framework based on the DeepSORT algorithm was deployed. 54 h of
high-resolution video data were collected at six signalized intersections (including three 3-leg and three 4-leg
intersections) in Glassboro, New Jersey. Non-compliance behaviors, such as redlight running and pedestrian
crossing outside the crosswalk, are captured to better understand the risky behaviors at these locations. The
proposed approach achieved an accuracy of 92% to 98% for detecting and tracking road users’ trajectories.
Additionally, the developed tool also provided directional traffic volumes, pedestrian volumes, vehicles running
a red light, pedestrian crossing outside the crosswalk events, and PET and TTC for crossing conflicts between
vehicles. Furthermore, an extreme value theory (EVT) was used to estimate the number of crashes at each
intersection utilizing the frequency of PETs and TTCs. Finally, the intersections were ranked based on the
calculated score considering the severity of crashes. Overall, the developed tool and the crash estimation, as well
as the model and ranking method, can provide valuable information for engineers and policymakers to assess the
safety of intersections and implement effective countermeasures to mitigate intersection-involved crashes.
1. Introduction collection process, the long time needed for data collection, and under-
reporting issues (Yang et al., 2021). Surrogate Safety Measures (SSMs)
In 2019, a total of 36,096 traffic fatalities occurred in the United have gained a great deal of attention as an alternative approach for road
States, according to the Federal Highway Administration (FHWA) safety analysis in recent years (Tarko et al., 2009; Zheng et al., 2014;
(FHWA, 2022). Approximately 10,180 of the total fatalities involved Laureshyn et al., 2017; Essa and Sayed, 2018). SSMs are also known as
intersections, suggesting a 1.6% increase compared to the fatalities that near-crash indicators, which measure and capture the temporal and
occurred in 2018. The great number of traffic fatalities and their in spatial proximity of road users. Traffic conflict is one of the proactive
crease in recent years has become a major concern for transportation SSM tools that can be implemented for road safety analysis. There are a
agencies, necessitating the importance of studying the safety of road variety of traffic conflict indicators being used for measuring interaction
users at intersections. Developing safety assessment analysis usually safety. Post-Encroachment Time (PET) and Time-to-Collision (TTC) are
needs a sufficient number of crash data. However, there are some lim two of the most commonly used traffic conflicts being applied for safety
itations in preparing crash data, including the difficulties in the analysis. High-definition trajectory data is critical for identifying traffic
* Corresponding author.
E-mail addresses: pateld80@rowan.edu (D. Patel), hossei93@rowan.edu (P. Hosseini), jalayer@rowan.edu (M. Jalayer).
1
ORCID: 0000-0001-7595-7489.
2
ORCID: 0000-0003-3488-3895.
3
ORCID: 0000-0001-6059-3942.
https://doi.org/10.1016/j.aap.2023.107264
Received 28 January 2023; Received in revised form 7 July 2023; Accepted 11 August 2023
Available online 4 September 2023
0001-4575/© 2023 Elsevier Ltd. All rights reserved.
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
conflicts, where several computer vision algorithms have been imple PET and TTC for vehicle conflicts, applying the EVT model to estimate
mented to recognize and track road users from video data (Simonnet crashes based on conflicts, and utilizing the Equivalent Property Damage
et al., 2012; Manh and Alaghband, 2018; Wang et al., 2019; Cai et al., Only (EPDO) score to rank the intersection based on the estimated crash
2019; Zhang et al., 2020). As part of this study, a real-time AI detection severity. This paper is organized as follows: the first part provides an
model - YOLO-v5, with a tracking framework based on the DeepSORT overview of prior research on the application of SSMs to intersection
algorithm, was implemented to extract the trajectory data. safety. The methodology section follows, where the data gathering
Once the trajectory data is extracted and traffic conflicts are identi approach and data analysis methods are described. The findings and
fied, they can be further analyzed by implementing different statistical their interpretation are detailed in the next part, while the conclusion
methods to develop a safety assessment of a specific road facility. In this and recommendations are offered in the last portion of this study.
study, the Extreme Value Theory (EVT) approach has been implemented
to estimate the possibility of extreme events from observations that 2. Literature review
occurred within a relatively short period. The EVT has been extensively
applied as a powerful tool in many research areas to assess the distri Examining intersection safety using SSMs has been implemented for
butions of extreme events. The EVT models are being proposed in safety evaluations at intersections in many studies over recent years. In
different approaches by using either one traffic conflict as a univariate 2015, Alhajyaseen created a crash hazard assessment that considers
model (Zheng et al., 2014; Zheng and Sayed, 2019a, 2019b) or more crash severity and occurrence probability to assess junction safety by
than one traffic conflict as a bivariate/multivariate model (Wang et al., using PET, collision angle, and total kinetic energy change before and
2019, Zheng and Sayed, 2019a; Zheng et al., 2018; Zheng et al., 2019b; after the impact as conflict indices. Results showed that the suggested
Zheng and Sayed, 2020). safety metric could successfully rank signalized junctions by collision
Considering the evolution made in traffic conflict methods, partic severity. In another study by Xie et al. (2016), safety improvements were
ularly those employing extreme value theory to establish the crash- assessed by using traffic conflicts obtained from 70 h of video recording.
conflict relationship, it has been demonstrated that traffic conflicts TTC for each vehicle pair was computed, and traffic control risks were
can effectively and reliably estimate both the overall frequency of determined afterward. As a result of this study, a robust correlation
crashes and crash frequency categorized by severity. However, many between traffic conflicts and actual crashes was observed. Chen et al.
studies have not yet concurrently assessed non-compliance behavior, (2017) used Unmanned Aerial Vehicle (UAV) footage to develop a sur
which remains one of the leading cause of crash fatalities. To address rogate safety analysis of pedestrian-to-vehicle conflicts at a Beijing
this issue and provide a comprehensive evaluation of intersection safety, intersection by using two SSMs, RTTC and PET. Results showed high
the following measures are employed: analyzing directional vehicle exposure of pedestrians to traffic conflict outside and inside of the
volumes, pedestrian volumes, events of vehicles running red lights, oc crosswalk. A study conducted in Cochabamba, Bolivia implemented an
currences of pedestrians crossing outside the crosswalks, calculating the automated video analytics tool named BriskLUMINA to assess SSMs and
2
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
3
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
4
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
Fig. 5. Positions of the violation bars (A), traffic lights (A), and crosswalk (B).
explained in the study conducted by Hou et al. (2019). counts and classifies the vehicle based on the direction flow was
developed. To obtain the flow direction of the vehicles, predefined zones
3.3. Conversion of pixel coordinates and speed estimation were created for each location. Each converted pixel value from the zone
was extracted and matched with the complete trajectories extracted
Estimating the position and speed of road users is a challenging from the model. Fig. 4 illustrates the zones that are created using the
problem in video analytics. The study applies the scale factor conversion OpenCV library and polygon plotting method. The flow direction of the
to address this problem by calculating the pixel per meter according to completely tracked objects was determined based on the start and
the camera perspective. The first frame is extracted from a given video endpoints of each track ID.
data and is divided into eight equal sections to identify the distribution
of road features and markings. Second, using the road features and 3.5. Traffic non-compliance counts
markings on the frame, several horizontal lines are drawn on the frames,
and the pixel distance between the start and endpoint of the line is Traffic non-compliance, such as vehicles running red lights and
calculated using the python OpenCV library. Third, the same lines are pedestrian crossing outside the crosswalk events, are crucial concerns at
drawn on Google Maps with reference to road features and markings, a signalized intersection. One element of the study was identifying ve
and the meter distance is evaluated. Lastly, the pixel per meter value is hicles running red lights. A predetermined traffic light region must be
calculated for each horizontal line by using Equation (1) as follows: chosen, and corresponding and adjacent violation bars must be drawn
Pd(x1, y1 ):(xn, yn ) on the frame. During the signal’s red phase, vehicles passing the cor
PixelPerMeter = (1) responding violation bar are considered vehicles running red light
Md(m) events. While during the signal’s green phase, vehicles passing the
adjacent violation bar are considered as vehicles running red light
where:
events. On traffic lights, the signal phase was determined by recognizing
Pd(x1, y1 ):(xn, yn ) : is the distance between the start and endpoint of the
the RGB values of the pixels that were predetermined by masking a
horizontal line drawn on the frame in coordinates. polygon using OpenCV library. For the situation where the intersection
Md(m) : is the distance between the start and endpoint of the line has a sign mentioning “NO TURN ON RED,” then the decision of the
drawn on the google map in meters. redlight running event can be configured by integrating the directional
It should be noted that an average pixel per meter value is considered traffic data. Fig. 5 (A) shows the position of the violation bars and traffic
for each section to convert the pixel coordinates. Fig. 3 demonstrates lights for the study locations.
eight sections on the frame and matching horizontal lines with the In terms of pedestrian crossing outside the crosswalk events, another
Google Map for one of the study locations. module of the analytics tool that functions using the python OpenCV
library and polygon plotting function was developed. A predefined re
3.4. Traffic counts gion is created on the frame covering all of the crosswalks and the
sidewalk areas. Then, a condition algorithm is deployed that continu
As the scope of this study, a system to identify the traffic flow that ously verifies the extracted trajectory data of the pedestrians. For
5
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
instance, if any coordinate values were observed outside the region, it For this study, TTC values that provided a negative time difference
was considered a crossing outside the crosswalk event. Fig. 5 (B) in were not considered for identifying the severity.
dicates the predefined polygon region created on the frame for the two
intersections.
3.7. Intersection safety analysis
the study also considered the 20 s as the arbitrary threshold for identi CR = Pr(D ≥ 0) = 1 − F(0) = 1 − exp{− 1+ξ } (5)
σ
fying all potential risks for vehicle-to-vehicle collisions at the
intersections. where:
CR: Crash risk between two vehicles
3.6.2. Time-to-collision (TTC) D: the negated PET or negated TTC
The collision time required for two vehicles if they keep driving at The calculated CR from equation (5) is a non-negative value. CR can
the same speed and the same path is defined as TTC (Hayward, 1972). As be considered as the crash frequency corresponding to the time period
discussed previously, the extracted trajectory, bounding box centroid, (t) from which the traffic conflicts are collected. By supposing that the t
bottom-front, and the bottom-back (i.e., the front and the rear of the is a representative of a period of interest denoting with T, the crash
vehicle) points were calculated and considered for the conflicting ve frequency that occurred during the T period can be computed as:
hicles. First, the intersecting point between the two trajectories was
identified, i.e., the conflict point. Then based on the conflicting point, N=
T
CR (6)
the vehicle that arrived first at the conflict point is assigned as the t
leading vehicle, and the vehicle that arrived second is considered as the
where:
following vehicle. After identifying the leading and the following vehi
N: number of crashes during the time period T.
cles, the trajectory data for the following vehicle is reconstructed to
In Equation (6), the T period is longer than the t period. It should be
reach the same conflict point following a similar path but at a different
mentioned that the number of extracted PET and TTC data for 3-leg
speed, which was observed before the deceleration occurred. With this
intersections was low. Hence, only 4-leg intersections were used to
process, the reconstructed trajectory will provide a projected arrival
develop the GEV models for yearly crash estimations.
time for the following vehicle, and then TTC can be determined using
By having the number of crashes during the time period T, a number
equation (3). A time–space diagram to calculate the TTC for vehicle-to-
of intersections (locations) can be ranked using different methods.
vehicle conflict is represented in Fig. 6.
Equivalent Property Damage Only (EPDO) is a commonly used crash
TTCt = tF,t˝ − tL,t (3) ranking method that is based on weighting by crash severities (Lim and
Kweon, 2013; Shams and Dissanayake, 2014; Manepalli and Bham,
where: 2016; Son et al., 2022). In this method, different weights are assigned to
tF,t˝ : the projected arrival time of the following vehicle at a conflict crashes with different severity levels in order to convert them equivalent
point after reconstructing the trajectory. to property damage-only crashes, known as EPDO score. The EPDO score
tL,t : the times the leading vehicle leaves at a conflict point. for each location can be calculated using the following equation:
6
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
7
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
Table 3
Detection results: Vehicle non-compliance counts (redlight running events) and pedestrian non-compliance counts (crossing outside the crosswalk).
Location Redlight Running Events Crossing Outside the Crosswalk
Total Vehicle Vehicle Non- Rate of Non- Total Pedestrian Pedestrian Non- Rate of Non-
Counts Compliance Counts Compliance Counts compliance Counts Compliance
Fig. 8. An illustration of a detected vehicle red-light running (A) and pedestrian crossing outside the crosswalk event (B). (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)
predefined zone discussed before. Fig. 7 illustrates the directional-based signal. However, this restriction may have been disregarded by the
count for one of the study locations. majority of drivers. Fig. 8 (A) visually depicts the detected vehicle non-
compliance event at the intersection of Donald Barger Blvd.
4.2. Traffic non-compliance counts
4.4. Pedestrians – Crossing outside the crosswalk
When understanding the safety of an intersection, non-compliance
behaviors of road users are imperative evaluation parameters. As part The aggregated results for all locations reveal that 1 in 4 pedestrians
of this study, the rate of non-compliance for the vehicles was examined crossing the street either do not completely use a crosswalk or cross
by extracting the redlight running events. In contrast, the rate of non- outside of designated crosswalk areas. Specifically, the intersections at
compliance for pedestrians was analyzed by extracting crossing High St. and Greentree Rd. exhibited the highest rates of non-
outside the crosswalk events for all junctions. compliance events, with values of 0.43% and 0.35% respectively.
Table 3 presents the results of the non-compliance rate of detected pe
destrians during a nine-hour period at each study location. Fig. 8 (B)
4.3. Vehicles – Redlight running events
visually depicts a detected pedestrian non-compliance event at the
intersection of High St.
The results of the detected red light running events for a duration of
nine hours at each study location are presented in Table 3. The in
tersections located at New St. and High St. exhibited the highest rates of 4.5. Surrogate safety measures
non-compliance, with values of 0.008% and 0.007% respectively.
Notably, neither of these locations had right turn on red signs installed, PET and TTC are SSMs used to identify the conflicts for left-turning
which means drivers were not allowed to make a right turn during a red vehicles. Nearly 120,000 complete trajectory data for all the locations
Table 4
Analysis results for PET and TTC.
Threshold Delsea Dr. N and Delsea Dr. N and Delsea Dr. N and Delsea Dr. N and Delsea Dr. N and Delsea Dr. N and Description
(Seconds) Donald Barger Blvd. Greentree Rd. Fishpond Rd. West St. New St. High St.
8
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
Table 5
Estimation results of GEV model and the calculated EPDO.
Model Location # of Blocks Estimate Standard Error Deviance Estimated Crashes (95% C.I.)
μ σ ξ μ σ ξ
GEV_PET Delsea Dr. N and High St. 32 − 1.245 0.976 − 0.782 0.192 0.194 0.200 71.326 0.157(0, 111.2)
Delsea Dr. N and New St. 56 − 1.419 0.927 − 0.727 0.134 0.121 0.110 122.302 0(0, 64.7)
Delsea Dr. N and Donald Barger Blvd. 37 − 1.273 0.727 − 0.552 0.136 0.117 0.166 70.839 0.584(0, 75.7)
GEV_TTC Delsea Dr. N and High St. 32 − 1.315 0.779 − 0.563 0.170 0.157 0.259 65.539 1.429(0, 95.4)
Delsea Dr. N and New St. 56 − 1.020 0.619 − 0.658 0.115 0.099 0.144 50.963 0(0, 74.8)
Delsea Dr. N and Donald Barger Blvd. 55 − 1.409 0.634 − 0.404 0.096 0.074 0.114 99.038 1.014(0, 42.8)
were extracted for accessing the PET and TTC. The trajectories with percentage of dangerous conflicts compared to other locations, with
conflicts of less than 20 s were considered to ease the further calcula 24.1% of the possible conflicts having a PET of less than 1.5 s. For the 4-
tions in this study. leg intersections, New St. demonstrated a higher percentage of
dangerous conflicts compared to other locations, with 12.1% of the
4.5.1. Post-encroachment time (PET) possible conflicts having a PET of less than 1.5 s. It is important to note
PET durations of less than 1.5 s indicate a higher likelihood of a crash that the frequency of conflicting events is heavily influenced by the
occurrence and represent dangerous conflicts (Zangenehpour et al., volume of left-turning traffic.
2015; Fu et al., 2016). Meanwhile, PET events occurring between 1.5
and 5 s suggest potential conflicts (Zangenehpour et al., 2015; Fu et al., 4.5.2. Time-to-collision (TTC)
2016). The PET results for each study location are presented in Table 4. Similar to PET, a TTC duration of less than 1.5 s indicates a higher
Among the 3-leg intersections, Fishpond Rd. exhibited a higher probability of a crash occurrence and represents a dangerous conflict. A
9
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
Fig. 10. Goodness of fit plots for TTC for three locations.
Fig. 11. Comparison between the observed crashes and the estimated crashes.
10
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
Table 6 Table 8
Severity ratios for the entire New Jersey. Calculated combined EPDO score and the ranking.
Severity Year Average Location Combined EPDO Score Ranking
2015 2016 2017 2018 2019 Delsea Dr. N and High St. 103.3 1
Delsea Dr. N and Donald Barger Blvd. 77.1 2
Fatal (K) 0.0027 0.0034 0.0027 0.0015 0.0035 0.0027
Delsea Dr. N and New St. 0.0 3
Suspected 0.0059 0.0038 0.0062 0.0056 0.0148 0.0072
Serious Injury
(A)
Suspected Minor 0.0629 0.0694 0.0683 0.0660 0.1338 0.0801
on their severity levels from 2015 to 2019. Thereafter, severity ratios for
Injury (B) different severity levels were calculated for each year separately.
Possible Injury 0.3119 0.3181 0.3237 0.3009 0.2300 0.2969 Finally, the ratios were averaged in order to get the yearly ratio for each
(C) severity for the entire study period. Table 6 tabulates the calculated
No Apparent 0.6166 0.6054 0.5992 0.6260 0.6179 0.6130
ratios for different severity levels for the entire state of New Jersey. As
Injury (O)
the next step, by having the severity distribution of the estimated
crashes for each location, EPDO crashes can be calculated. It is note
TTC event occurring between 1.5 and 5 s suggests a potential conflict. worthy to remark that the EPDO index was computed using calibrated
The calculated TTC results for each study location are displayed in ratios according to the New Jersey crash costs provided by the 2016
Table 4. Consistent with the PET findings, Fishpond Rd., among the 3-leg Highway Safety Manual (Harmon et al., 2018).
intersections, and New St., among the 4-leg intersections, exhibited a The calculated severity distribution of the estimated crashes, as well
higher rate of risky conflicts. as the EPDO indices, are listed in Table 7. As can be seen in this table, the
highest number of fatal crashes was estimated for Delsea Dr. N and High
4.6. Intersection safety modeling St. intersection for GEV model with TTC as the traffic conflict indicator.
Moreover, the EPDO index for this location was computed as 11.6.
In order to model the crash risk, a BM approach was implemented in Subsequently, the EPDO indices were combined to get the final
which the traffic conflict indicator (e.g., PET or TTC) is divided into values for ranking the intersections. According to a previous study
fixed time intervals. According to the previous studies, generally, more (Zheng and Sayed, 2019a), on average, the estimated crashes from the
than 30-time intervals for taking the block maxima are required in order GEV model using TTC were around nine times the estimated crashes
to fit a model (Zheng et al., 2014; Zheng and Sayed, 2019a. In this study, from the GEV model using PET. This relationship between the crashes
GEV models with four different block intervals were developed for the estimated from TTC and PET models was implemented for weighting the
three locations (4-leg intersections). Table 5 tabulates the selected EPDO indices for calculating the final combined EPDO scores. The
number of blocks and the model estimation results for each location. By combined EPDO score can be computed by summation of the weighted
substituting the estimated parameters from the developed models in EPDOs from PET and TTC models for each location. Finally, the three
Equation (5), the number of yearly crashes for the three locations was intersections were ranked based on their combined EPDO scores. The
estimated. Moreover, a 95% confidence interval (CI) of the estimated lower the value of the combined EPDO score, the safer the intersection
crashes was calculated. Table 5 also lists the estimated crashes and their would be. Table 8 presents the computed combined EPDO and the
95% CI. rankings for the three 4-leg intersections. As can be seen in this table,
Figs. 9 and 10 illustrate the goodness of fit plots for the developed Delsea Dr. N and High St. intersection (EPDO Score:103.3) was found to
models using PET and TTC. According to the QQ plots, the difference be the un-safest intersection in terms of safety, followed by the Delsea
between empirical and model values is very small for both PET and TTC Dr. N and Donald Barger Blvd intersection (EPDO Score:77.1). More
results. The Density plots also indicate that the models developed using over, Delsea Dr. N and New St intersection (EPDO Score:0.0) was ranked
PET and TTC fitted well. as the safest intersection among other locations.
Fig. 11 illustrates the comparison between the estimated crashes by
the models and the observed crashes. As shown in this figure, overall, 5. Conclusion
the left-turn crashes by the models with PET and TTC were both esti
mated close to the observed left-turn crashes for all locations. Only the The main intention of this study was to proactively evaluate the
crashes using TTC for Delsea Dr. N and Donald Barger Blvd. intersections safety of intersections by using traffic conflicts and non-compliance
were estimated far from the observed crashes. Moreover, the 95% CI. of behavior. To achieve this study’s goal, an innovative AI and a web-
the estimated crashes using TTC were narrower than those of the esti based Safety Analytics application were developed to detect, track,
mated crashes using PET. count, and recognize non-compliance behavior and identify road user
The final step for evaluating the safety of intersections is to rank the conflicts and conflict severity by SSMs. In this study, PET and TTC were
three locations in terms of their safety. In this study, the intersections are used to identify future threats associated with left-turning vehicles. Af
ranked based on the value of their estimated EPDO crashes. The esti terward, nine hours of video data were recorded from six intersections
mated number of left-turn crashes from GEV models with PET and TTC located in Glassboro, New Jersey, to evaluate the performance of the
are total left-turn crashes. In order to get the severity distribution of the developed tools, as well as to assess intersection safety. Overall, 54 h of
crashes, the New Jersey severity ratios were used. To calculate the high-resolution video data consisting of more than 5.8 million frames
severity ratios, the left-turn crashes for New Jersey were obtained based were processed using the tool that integrated a real-time AI detection
Table 7
Calculated severity distribution of the estimated crashes.
Model Location Estimated Crashes K A B C O EPDO
GEV_PET Delsea Dr. N and High St. 0.157 0.0004 0.0011 0.0126 0.0466 0.0963 1.3
Delsea Dr. N and New St. 0 0.0000 0.0000 0.0000 0.0000 0.0000 0.0
Delsea Dr. N and Donald Barger Blvd. 0.584 0.0016 0.0042 0.0468 0.1735 0.3582 4.7
GEV_TTC Delsea Dr. N and High St. 1.429 0.0039 0.0103 0.1144 0.4242 0.8758 11.6
Delsea Dr. N and New St. 0 0.0000 0.0000 0.0000 0.0000 0.0000 0.0
Delsea Dr. N and Donald Barger Blvd. 1.01 0.0028 0.0073 0.0812 0.3010 0.6214 8.2
11
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
model - YOLO-v5 with a tracking framework based on the DeepSORT The State University or those of the U.S. Department of Transportation, Of
algorithm to extract trajectory data. The integrated model demonstrated fice of the Assistant Secretary for Research and Technology (OST-R).”.
an accuracy between 92 and 99%. The results showed that New St. and
High St. intersection had the highest rate of non-compliance for vehicles References
running a red light. Regarding the pedestrian non-compliance events, it
was observed that 1 in 4 pedestrians crossing the street are not entirely Alhajyaseen, W.K., 2015. The integration of conflict probability and severity for the
safety assessment of intersections. Arab. J. Sci. Eng. 40 (2), 421–430.
using a crosswalk or are crossing outside the crosswalk. For the conflict Amundsen, F.H., Hyden, C., 1977. Proceedings of the 1st Workshop on Traffic Conflicts,
analysis, PET and TTC results showed that the intersections at Fishpond Oslo, Norway.
Rd. and New St. had the highest percentage of dangerous conflicts Arun, A., Haque, M.M., Bhaskar, A., Washington, S., 2022. Transferability of multivariate
extreme value models for safety assessment by applying artificial intelligence-based
among other 3-leg and 4-leg intersections, respectively. video analytics. Accid. Anal. Prev. 170, 106644.
As the next step of this study, by using the extracted PETs and TTCs Bewley, A., Ge, Z., Ott, L., Ramos, F., & Upcroft, B. (2016, September). Simple online and
from the video data, GEV models with BM approach were developed to realtime tracking. In 2016 IEEE international conference on image processing (ICIP)
(pp. 3464-3468). IEEE.
estimate the number of yearly left-turn crashes for the 4-leg in Cai, Q., Abdel-Aty, M., Sun, Y., Lee, J., Yuan, J., 2019. Applying a deep learning
tersections. Eventually, by calculating the EPDO scores, the 4-leg in approach for transportation safety planning by using high-resolution transportation
tersections were ranked regarding their safety. The ranking results and land use data. Transp. Res. A Policy Pract. 127, 71–85.
Chen, P., Zeng, W., Yu, G., Wang, Y., 2017. Surrogate Safety Analysis of Pedestrian-
revealed that Delsea Dr. N and West St. intersection is the unsafe loca
Vehicle Conflict at Intersections Using Unmanned Aerial Vehicle Videos. J. Adv.
tion among all the 4-log intersections. Transp. 2017, 1–12.
As a limitation of this paper, it is important to note that the crash Cooper, D.F., Ferguson, N., 1976. Traffic studies at T-Junctions. 2. A conflict simulation
estimation model does not take into account the unobserved heteroge Record. Traffic Eng. Control 17 (Analytic).
Essa, M., Sayed, T., 2018. Full Bayesian conflict-based models for real time safety
neity among the intersections. This means that the unique characteris evaluation of signalized intersections. Accid. Anal. Prev. 129, 367–381.
tics of each intersection, which could potentially affect the likelihood of Federal Highway Administration (FHWA). (2022). About Intersection Safety, U.S.
crashes occurring, are not fully accounted for in the model. Hence, as a Department of Transportation (USDOT). https://safety.fhwa.dot.gov/intersection/
about/.
potential avenue for future research, this limitation could be addressed Fu, T., Miranda-Moreno, L., Saunier, N., 2016. Measuring crosswalk safety at
by incorporating random effects or coefficient approaches into the crash nonsignalized crossings during nighttime based on surrogate measures of safety:
estimation model, and integrating non-compliance events into the Case study in Montreal, Canada. In: Transportation Research Board Annual Meeting
Compendium of Papers.
evaluation process would contribute to a more comprehensive assess Fu, C., Sayed, T., Zheng, L., 2020. Multivariate Bayesian hierarchical modeling of the
ment. As another part of future research, more video footage from non-stationary traffic conflict extremes for crash estimation. Anal. Methods Acc. Res.
different sources, such as 511.org, PTZ cameras, and other live streams, 28, 100135.
Fu, C., Sayed, T., 2022. A multivariate method for evaluating safety from conflict
will be utilized to assess the model’s robustness. Other safety indicators, extremes in real time. Anal. Methods Acc. Res. 36, 100244.
like Deceleration Rate (DR), Gap Time (GT), Proportion of Stopping Harmon, T., Bahar, G.B., Gross, F.B., 2018. Crash costs for highway safety analysis (No.
Distance (PSD), and the severity of the non-compliance event could be FHWA-SA-17-071). United States. Federal Highway Administration. Office of Safety.
Hayward, J.C., 1972. Near miss determination through use of a scale of danger.
determined to assess intersection safety extensively. Moreover, a longer
Hou, X., Wang, Y., & Chau, L. P. (2019, September). Vehicle tracking using deep sort
duration of video data will be recorded for 3-leg intersections in order to with low confidence track filtering. In 2019 16th IEEE International Conference on
include them in the analysis. Lastly, bivariate extreme value models will Advanced Video and Signal Based Surveillance (AVSS) (pp. 1-6). IEEE.
also be developed to compare the crash estimation results with the Hussain, F., Li, Y., Arun, A., Haque, M.M., 2022. A hybrid modelling framework of
machine learning and extreme value theory for crash risk estimation using traffic
univariate models developed in this study. conflicts. Anal. Methods Acc. Res. 36, 100248.
Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., NanoCode012, Kwon, Y., TaoXie, Fang,
Authors contribution J., Imyhxy, Michael, K., Lorna, V, A., Montes, D., Nadar, J., Laughing, Tkianai,
yxNONG, Skalski, P., Wang, Z., Minh, M.T., 2022, February 22. Ultralytics/yolov5:
V6.1 - tensorrt, tensorflow edge TPU and OpenVINO export and inference. Zenodo.
The authors confirm contribution to the paper as follows: Deep Retrieved July 2, 2022, from https://zenodo.org/record/6222936#.YsAItXbMKUk.
Patel: study conception and design, data collection, analysis and Laureshyn, A., de Goede, M., Saunier, N., Fyhri, A., 2017. Cross-comparison of three
surrogate safety methods to diagnose cyclist safety problems at intersections in
interpretation of results, draft manuscript preparation. Parisa Hosseini: Norway. Accid. Anal. Prev. 105, 11–20.
analysis and interpretation of results, draft manuscript preparation. Lim, I.K., Kweon, Y.J., 2013. Identifying high-crash-risk intersections: comparison of
Mohammad Jalayer: study conception and design, draft manuscript traditional methods with the empirical bayes-safety performance function method.
Transp. Res. Rec. 2364 (1), 44–50.
preparation. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Zitnick, C.L., 2014.
In: Microsoft Coco: Common Objects in Context European Conference on Computer
Declaration of Competing Interest Vision. Springer, Cham, pp. 740–755.
Manepalli, U.R.R., Bham, G.H., 2016. An evaluation of performance measures for hotspot
identification. J. Transp. Safety Security 8 (4), 327–345.
The authors declare that they have no known competing financial Manh, H., Alaghband, G., 2018. Scene-lstm: A model for human trajectory prediction.
interests or personal relationships that could have appeared to influence arXiv preprint arXiv:1808.04018.
Redmon, J., Divvala, S., Girshick, R., Farhadi, A., 2016. You only look once: Unified,
the work reported in this paper.
real-time object detection. In: Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 779–788.
Data availability Scholl, L., Elagaty, M., Ledezma-Navarro, B., Zamora, E., Miranda-Moreno, L., 2019.
A surrogate video-based safety methodology for diagnosis and evaluation of low-cost
pedestrian-safety countermeasures: the case of cochabamba, bolivia. Sustainability
Data will be made available on request. 11 (17), 4737.
Shams, A., Dissanayake, S., 2014. Improving safety at unsignalized rural intersections in
Acknowledgments Kansas. In: T&DI congress 2014: Planes, trains, and automobiles, pp. 356–365.
Simonnet, D., Velastin, S.A., Turkbeyler, E., Orwell, J., 2012. Backgroundless detection
of pedestrians in cluttered conditions based on monocular images: a review. IET
“This publication was supported by a subaward from Rutgers University, Comput. Vis. 6 (6), 540–550.
United States, Center for Advanced Infrastructure & Transportation, under Son, S.O., Park, J., Lee, G., Abdel-Aty, M., 2022. Development of new performance
measures based on data mining weights for hotspot identification. Transp. Res. Rec.
Grant no. 69A3551847102 from the U.S. Department of Transportation, 2676 (8), 633–647.
Office of the Assistant Secretary for Research and Technology (OST-R).” All Tarko, A., Davis, G., Saunier, N., Sayed, T., Washington, S., 2009. Surrogate Measures of
materials, except scientific articles or papers published in scientific Safety. Transportation Research Board, Washington, DC. White paper.
Wang, C., Xu, C., Xia, J., Qian, Z., Lu, L., 2018. A combined use of microscopic traffic
journals, must also contain the following disclaimer: “Any opinions, simulation and extreme value methods for traffic safety evaluation. Transp. Res. Part
findings, and conclusions or recommendations expressed in this publication C: Emerg. Technol. 90, 281–291.
are those of the author(s) and do not necessarily reflect the views of Rutgers,
12
D. Patel et al. Accident Analysis and Prevention 192 (2023) 107264
Wang, C., Xu, C., Dai, Y., 2019. A crash prediction method based on bivariate extreme Zheng, L., Sayed, T., 2019a. From univariate to bivariate extreme value models:
value theory and video-based vehicle trajectory data. Accid. Anal. Prev. 123, approaches to integrate traffic conflict indicators for crash estimation. Transp. Res.
365–373. Part C: Emerg. Technol. 103, 211–225.
Xie, K., Li, C., Ozbay, K., Dobler, G., Yang, H., Chiang, A. T., & Ghandehari, M. (2016, Zheng, L., Sayed, T., 2019b. Bayesian hierarchical modeling of traffic conflict extremes
November). Development of a comprehensive framework for video-based safety for crash estimation: a non-stationary peak over threshold approach. Anal. Methods
assessment. In 2016 IEEE 19th International Conference on Intelligent Transportation Acc. Res. 24, 100106.
Systems (ITSC) (pp. 2638-2643). IEEE. Zheng, L., Sayed, T., 2020. A bivariate Bayesian hierarchical extreme value model for
Yang, D.i., Xie, K., Ozbay, K., Yang, H., 2021. Fusing crash data and surrogate safety traffic conflict-based crash estimation. Anal. Methods Acc. Res. 25, 100111.
measures for safety assessment: Development of a structural equation model with Zheng, L., Ismail, K., Meng, X., 2014. Freeway safety estimation using extreme value
conditional autoregressive spatial effect and random parameters. Accid. Anal. Prev. theory approaches: a comparative study. Accid. Anal. Prev. 62, 32–41.
152, 105971. Zheng, L., Ismail, K., Sayed, T., Fatema, T., 2018. Bivariate extreme value modeling for
Zangenehpour, S., Miranda-Moreno, L.F., Saunier, N., 2015. Automated classification road safety estimation. Accid. Anal. Prev. 120, 83–91.
based on video data at intersections with heavy pedestrian and bicycle traffic: Zheng, L., Sayed, T., Essa, M., 2019a. Validating the bivariate extreme value modeling
Methodology and application. Transp. Res. Part C: Emerg. Technol. 56, 161–176. approach for road safety estimation with different traffic conflict indicators. Accid.
Zhang, S., Abdel-Aty, M., Wu, Y., Zheng, O.u., 2020. Modeling pedestrians’ near-accident Anal. Prev. 123, 314–323.
events at signalized intersections using gated recurrent unit (GRU). Accid. Anal. Zheng, L., Sayed, T., Essa, M., 2019b. Bayesian hierarchical modeling of the non-
Prev. 148, 105844. stationary traffic conflict extremes for crash estimation. Anal. Methods Acc. Res. 23,
100100.
13