Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye

Video

Rahul Sengupta

, Tania Banerjee

, Ke Chen, Yashaswi Karnati

, Sanjay Ranka

and Anand Rangarajan

Department of Computer and Information Science & Engineering, University of Florida, Gainesville, FL, U.S.A.

ﬂ

Keywords:

Simulation and Modeling, Data Mining and Data Analysis, Modeling, Simulation, Control of Pedestrians and

Cyclists.

Abstract:

Most trafﬁc authorities across the US usually collect high-resolution (10 Hz) loop detector and signal state

data and video data. The multiple modalities of data that are readily available can be utilized for better trafﬁc

operations management and improving safety. In this work, we show that the fusion of widely deployed loop

detector data with trajectory information collected through video cameras can augment intersection safety and

operational efﬁciency analysis. The additional information that can be extracted from the object’s (vehicle and

pedestrian) trajectory derived from video data when fused with signal state data leads to several interesting

safety analyses. Data analysis shows a signiﬁcant variance in turn-movement counts, pedestrian behaviors,

vehicle composition, etc., temporally (hour-of-day, day-of-week, etc.) and spatially (approach-wise). We

present a simulation-based approach for customizing signal timing plans based on the trafﬁc behavior at the

intersections at various times. When used to drive simulations in demand generation, we show that the fused

data calibrating the simulation parameters can lead to potential improvements in existing signal timing plans

that match reality and can greatly help improve intersection safety and operational efﬁciency by providing

planners with data-driven insights.

1 INTRODUCTION

With rapid urbanization ocurring worldwide, the

growing volume of vehicles and increasing complex-

ity of road networks have led to problems such as con-

gestion, trafﬁc jams, and trafﬁc incidents (Rao and

Rao, 2012; Carson et al., 2010). These have been

shown to affect productivity in a negative way and

also the local economy(Weisbrod et al., 2003), soci-

etal well-being (Levy et al., 2010), and the environ-

ment (Zhang and Batterman, 2013). Therefore, the

smooth ﬂow of trafﬁc and safety are important con-

cerns for trafﬁc authorities.

Intelligent Transportation Systems is (ITS) a fast-

growing ﬁeld (Alam et al., 2016; Borgi et al., 2017)

and some commonly seen aspects of ITS (Gordon,

2016) include:

https://orcid.org/0000-0001-9793-5176

https://orcid.org/0000-0003-4737-0001

https://orcid.org/0000-0002-2512-1250

https://orcid.org/0000-0003-4886-1988

https://orcid.org/0000-0001-8695-8436

• Use of microprocessor-based trafﬁc signals that

are “coordinated” to optimize trafﬁc ﬂow across

a corridor. These use embedded road sensors to

relay vehicle detection counts

• Collection and storage of high-resolution (10 Hz)

induction loop detector actuations along with de-

tailed signal state information, as well as visual-

ization of derived metrics called Automated Traf-

ﬁc Signal Performance Measures

• Use of video cameras (with computer vision pro-

cessing) at intersections to detect trafﬁc signal vi-

olations, vehicle tracking, queue length estima-

tion, etc.

Most trafﬁc authorities across the United States

of America usually collect high-resolution (10 Hz)

loop detectors and signal state data (Figure 1). How-

ever, there are signiﬁcant drawbacks, such as the in-

ability to collect precise trajectory information, turn-

movement counts, pedestrian walking behaviors, etc.

We focus on using data obtained from ﬁsheye lens

Link: atspm.cﬂsmartroads.com/ATSPM/

Sengupta, R., Banerjee, T., Chen, K., Karnati, Y., Ranka, S. and Rangarajan, A.

Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye Video.

DOI: 10.5220/0011729000003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 54-63

ISBN: 978-989-758-652-1; ISSN: 2184-495X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Figure 1: A sample of high-resolution loop detector and

signal state data.

Figure 2: Table showing the beneﬁts of using ﬁsheye lens

video data compared to the pre-existing high-resolution

loop detector and signal state data.

cameras, which can mitigate some of these deﬁcien-

cies.

As shown in Figure 2, items ‘ a’ and ‘b’ help com-

pute descriptive statistics of the overall intersection

dynamics. Items ‘c’ and ‘d’ are vital for intersection

safety, whereas ‘a’,‘b’, ‘c’, ‘d’, ‘e’, and ‘f’ are es-

sential for building an accurate, calibrated simulation

model for signal timing optimization.

The collection, processing, and application of in-

formation extracted from ﬁsheye video data can be

effectively augmented with high-resolution loop de-

tectors and signal state data for applications in safety

analyses and improving trafﬁc operations. The ad-

ditional information that can be extracted from a

user’s (vehicle or pedestrian) trajectory, such as speed

proﬁles, vehicle-to-vehicle interactions, vehicle-to-

pedestrian interactions, abrupt braking, etc., can be

helpful to qualitatively and quantitatively understand

intersection safety. Also, this information when com-

bined with signal state data can lead to further inter-

esting analysis, such as red-light violations, vehicle

movements during red clearance intervals, etc.

The use of microscopic simulation to understand

intersection dynamics, including the impact of differ-

ent signal timing plans, is quite common. In order

to ensure that the trafﬁc ﬂow dynamics in a simu-

lator align with real-world observations, incorporat-

ing real-world data into simulations for demand gen-

eration and calibration of simulation parameters is a

crucial aspect. The controller log data can be used

to obtain granular vehicle ﬂow data in the simula-

tor. However, video data can provide accurate turning

movement counts (especially for combined turn lanes

such as through-right turning lanes), vehicle composi-

tions, speed/acceleration proﬁles, accurate pedestrian

demand, and walking speeds. These can be used to

ﬁne-tune and calibrate the simulation and thus make

it more realistic.

Using data collected from real trafﬁc intersections,

we show that potential improvements in the existing

signal timing can be made using controller log data

and additional information from video data.

We have deployed a ﬁsheye video collection sys-

tem at a vital intersection in a dense urban region

with time-varying vehicular and pedestrian trafﬁc. We

have processed the video obtained and extracted tra-

jectory information for vehicles and pedestrians. Our

data analysis shows a signiﬁcant variance in turn-

movement counts, pedestrian behaviors, vehicle com-

position, etc., temporally (hour-of-day, day-of-week,

etc.) and spatially (approach-wise). Such analysis can

signiﬁcantly help improve intersection safety and op-

erational efﬁciency by providing planners with data-

driven insights. The overall system ﬂowchart is pre-

sented in Figure 3. The main contributions of the pa-

per are as follows:

• We quantitatively and qualitatively present safety

analyses at an intersection by fusing controller log

data with trajectory information obtained through

video cameras.

• We present a simulation-based approach for cus-

tomizing signal timing plans at an intersection at

various times. When used to drive simulations,

the fused data can lead to potential improvements

in existing signal timing plans that match reality.

The rest of the paper is organized as follows. Sec-

tion 2 discusses the background information. Sec-

tion 3 presents the details of our ﬁeld implementation.

We also review the data obtained and present some

descriptive statistics. We describe the application of

the data obtained to intersection safety in Section 4

and signal timing optimization in Section 5. We con-

clude and present future work in Section 6.

2 BACKGROUND

We now discuss the data collection methodology and

the simulation software.

Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye Video

Figure 3: Flowchart showing the workﬂow for using ﬁsheye video data to improve safety and signal timing of an intersection

using a trafﬁc microsimulation framework. Fisheye video data provides vital information on different aspects of the intersec-

tion dynamics, which can then be used to improve modeling.

2.1 Fisheye Lens Camera Data:

Collection and Processing

A ﬁsheye lens is a type of camera lens that produces

a distortion that allows for an ultra-wide-angle ﬁeld

of view, often 100 to 180 degrees. This allows a sin-

gle camera to capture a much larger scene than regu-

lar camera lenses. Such lenses are often paired with

video cameras to record the dynamics of trafﬁc inter-

sections. Additional video processing techniques are

required to perform object-tracking in such distorted

videos.

While computer vision using regular non-

distorted video data at trafﬁc intersections is well-

explored (Mondal et al., 2019; Buch et al., 2011; San-

thosh et al., 2020), there needs to be more research

using ﬁsheye video data. (Huang et al., 2020) uses

an integrated two-stream convolutional networks ar-

chitecture that performs real-time detection, tracking,

and near-accident detection of different road users

(pedestrains and vehicles) in trafﬁc video from a ﬁsh-

eye lens camera; (Yeh et al., 2020) details a sys-

tem for automatically identifying and tracking vehi-

cles and processing their trajectories; (Chen et al.,

2021) describes a visualization tool for analyzing tra-

jectories from ﬁsheye video data; (Zhao et al., 2021)

discusses a system for pedestrian detection and re-

identiﬁcation.

In our video processing pipeline based on (Chen

et al., 2021; Huang et al., 2020), raw video data from

the ﬁsheye lens was processed, and trajectory infor-

mation of vehicles and pedestrians was obtained. Ob-

ject detection was done using YOLOV4(Bochkovskiy

et al., 2020) Deep Neural Network, and multiple-

object tracking was built upon Deep SORT(Wojke

et al., 2017). Camera calibration was performed, and

landmark points were mapped from the ﬁsheye cam-

era image of the intersection to the top-view satel-

lite map image of the intersection. Thin-Plate Spline

(Bookstein, 1989; Chui and Rangarajan, 2000) was

used as the basis function for coordinate mappings

from the reference to the target. Thus, object trajecto-

ries seen in the ﬁsheye videos were projected onto the

satellite map of the intersection. Outlier detection and

smoothing techniques were used to stabilize trajecto-

ries. Manual annotation and checking were also per-

formed to get a ﬁnal accuracy of over 95 percent for

detecting vehicles and pedestrians. The details of the

performance of the vision processing algorithms can

be found in (Chen et al., 2021; Huang et al., 2020).

2.2 Trafﬁc Simulation Software

Trafﬁc simulation frameworks (US DOT FHWA,

2020; DLR, 2020) are computational implementa-

tions of trafﬁc models with dynamic components (ve-

hicles, trafﬁc signals, pedestrians, etc.) and static

components (road geometry and linkages, etc.). A

trafﬁc simulation consists of a trafﬁc scenario with

a base map that deﬁnes the static features, such as

the topology of roads with lanes, junctions that con-

nect these roads, etc. These static components usu-

ally remain the same in the short term (i.e., in sec-

onds or minutes). On this base map, dynamic com-

ponents usually change their states based on prede-

ﬁned behaviors (i.e., cars will change their locations

based on their speeds and accelerations, trafﬁc sig-

nals will change their light conﬁguration based on the

signal plan, etc.). The simulation is started and is al-

lowed to evolve in time. We can thus simulate a vari-

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

ety of base maps and behaviors and estimate different

measures of effectiveness (such as queue lengths and

travel times).

One such essential simulation software is Sim-

ulation of Urban MObility (SUMO) (Lopez et al.,

2018). SUMO is an open-source, microscopic, agent-

based road trafﬁc simulation package that is de-

signed to handle large road networks. SUMO uses

its ﬁle formats for trafﬁc networks, but it can im-

port ﬁles encoded in other popular formats like Open-

StreetMap (OpenStreetMap contributors, 2017), VIS-

SIM (Lownes and Machemehl, 2006) etc. SUMO is

implemented in C++ and uses only portable libraries,

thus making it lightweight and fast. SUMO is single-

threaded, i.e., it uses only one CPU core, but several

parallel SUMO processes can be spawned, allowing

for parallel simulations.

3 FIELD IMPLEMENTATION

AND DESCRIPTIVE

STATISTICS

A critical intersection at a major urban center in

the United States of America was analyzed for this

demonstration. The intersection borders a large uni-

versity. Several residential complexes and commer-

cial establishments (especially restaurants) in the area

give rise to signiﬁcant pedestrian trafﬁc and vehicular

trafﬁc. On the median, a single ﬁsheye video camera

is mounted 13 feet (3.96 meters) above the intersec-

tion at the southbound approach.

Raw data from the intersection (ﬁsheye video

data, high-resolution loop detector, and signal state

data) was captured and stored in a cloud database.

Three weeks of data, spanning October-November

2021, with time ranges from 6 am to 8 pm, was con-

sidered. This time range captures the AM (morning),

PM (evening), and midday trafﬁc peaks but also en-

sures sufﬁcient ambient light for the camera to func-

tion. The Major ﬂow (i.e., North-South direction)

usually sees higher ﬂows than the Minor ﬂow (i.e.,

East-West direction), with the Major ﬂow (North-

South combined) reaching 1400-1800 vph (vehicles

per hour) at peak and Minor ﬂow (East-West com-

bined) reaching 800-1200 vph.

We then analyze the data given the hour-of-day

and day-of-week. We look for both variations and

similarities. These will help us identify periods where

signiﬁcant pedestrian trafﬁc co-exists with vehicu-

lar trafﬁc and also inform us where segmented cus-

tomized signal timing plans may be needed.

4 APPLICATION TO IMPROVE

INTERSECTION SAFETY

Fisheye cameras provide a crucial piece of informa-

tion: turn-movement counts. This data is challeng-

ing to obtain from loop-detector-based ATSPM data

without exit detectors. Intersections frequently have

lanes that accommodate multiple movements, such as

the right-most lane allowing through and right-turning

trafﬁc. Fisheye lens tracking can count these vehi-

cles, which loop detectors cannot differentiate. Fish-

eye data is necessary to analyze direction-wise ﬂows

(see Figure 4). In addition, ﬁsheye cameras also al-

low us to track individual pedestrians. This allows us

to estimate the number of pedestrians crossing in var-

ious directions. This contrasts with ATSPM data that

tracks the number of pedestrian calls but has no way

of estimating the actual number of pedestrians (Figure

5, Figure 7). With accurate trajectories of vehicles

and pedestrians, it is possible to study, both qualita-

tively and quantitatively, the safety aspects of the in-

tersection. Visualizations (Figure 4, Figure 5, Figure

7) can help determine overlapping times and direc-

tions where vehicles and pedestrians have intersect-

ing trajectories. This can support trafﬁc authorities

in modifying intersection signaling (e.g., restricting

right turns, etc.) for those times and directions.

Standard safety metrics(Mullakkal Babu et al.,

2017) such as Time-To-Collision (TTC) and Post-

Encroachment Time (PET) can also be calculated.

Our intersection safety research (Banerjee et al.,

2022) has involved computing TTC and PET, along

with other factors such as speed, deceleration, prox-

imity between conﬂicting users, to determine severe

events. These metrics are solely obtainable from

trajectory datasets that are processed through video.

In (Mishra et al., 2022), we presented a study that

showed how the day of the week and time of day can

affect the frequency and severity of events at intersec-

tions. This analysis enables us to pinpoint the busiest

periods and conﬂict-prone locations at intersections.

The data can then be utilized to evaluate safety coun-

termeasures with minimal disruption to intersection

efﬁciency.

The ﬁsheye video also lets us detect heavy motor

vehicles, i.e., large vehicles such as delivery trucks

and buses (Figure 6). These large vehicles often

have a long braking distance and signiﬁcant blind

spots. Given that this intersection borders the Uni-

versity and has several shopping and eating locations

nearby, these large vehicles may pose a safety concern

for pedestrians. Their movement can be restricted or

rerouted based on such data. Using this information

to adjust signal timing plans for transit priority is also

Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye Video

Figure 4: Heatmap showing the intensity of vehicles per hour for various hours of the day across the week. We can see

that the Southbound-through direction has a larger AM peak, and the Northbound-through direction sees a larger PM peak.

This is likely because the University is south of the intersection, with people coming to work in the morning and leaving

in the evening. Similarly, for Southbound-Right and Northbound-Right directions, we can see that both AM and PM peaks

for Southbound-Right are larger. Loop detectors do not usually capture right turns as they share the same lane as through

movements. Right turns are vital for pedestrian safety, as drivers may neglect pedestrians while focusing on merging at the

right turn.

possible. Thus, we have seen the usefulness of ﬁsh-

eye lens camera videos in providing us with statis-

tics and visualizations of various road users, includ-

ing heavy vehicles and pedestrians. Such analysis can

lead to better policy decisions, especially for pedes-

trian safety.

5 APPLICATION TO SIGNAL

TIMING OPTIMIZATION

In this section, we present our efforts in utilizing the

ﬁsheye lens camera data for modeling the intersection

dynamics in simulation. We can then vary various sig-

nal timing parameters to improve performance.

We build a simulation (Figure 9) of the chosen in-

tersection. We then input ﬂows by the cycle with the

correct turn-movement counts. We also replicate the

effect of pedestrian calls within the ring-and-barrier

actuated signal timing plan we have implemented in

the simulation. Using bounding box information, we

can further calibrate vehicle lengths and types in the

simulation.

Signiﬁcant variations in different aspects of traf-

ﬁc patterns will indicate to us that customized signal

timing plans may be required. The similarities across

such sections suggest which signal timing plan can be

effectively applied in the future. We plot and analyze

the weekly trafﬁc throughput in Figure 10. This leads

us to identify four crucial stable high-volume trafﬁc

patterns:

• Weekday AM peak,

• Weekday midday peak

• Weekday PM peak

• Weekend PM peak

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

Figure 5: Heatmap showing the intensity of pedestrians per hour for various hours of the day, across three weeks along the

Minor Phase (i.e., East-West direction). We see high activity around lunch hours on weekdays. An unusual peak was seen on

Saturday in the second week of November. This can be explained due to an important football game that day. Such anomalies

can be inferred from this data.

Figure 6: Heatmap showing the intensity of heavy vehicles

per hour for various hours of the day across the week. We

can see that heavy vehicles are usually limited to weekday

mornings. Additional measures, such as restricting right-

turns, re-routing, etc., can be put into effect at those times

for safety reasons.

Figure 7: Heatmap showing the number of pedestrian calls

per hour for various hours of the day across three weeks

along the Major Phase (i.e., North-South direction). We see

high activity around lunch hours on weekdays, where al-

most every cycle has a pedestrian call. Customized signal

behavior (turn restrictions etc.) and timing plans can be im-

plemented.

We simulate these scenarios with their respective

signal timing plans, vehicle ﬂow patterns, and pedes-

trian call patterns. We vary the barrier times (sepa-

rating the major and minor phase rings in the Ring-

and-Barrier signal plan) across a reasonable range of

values and see the effect on the trade-off between the

major and minor phase wait times.

We ﬁrst see the impact of introducing pedestrian

calls (Figure 8) and ignoring them. We can see an

apparent effect of including pedestrian calls. Pedes-

trians require signiﬁcant time to cross the road (here,

30 seconds) safely. The ring-and-barrier conﬁgura-

tion shown, including pedestrian calls within a cycle,

ensures lower time for the left-turning phases that pre-

cede them (in both major and minor directions). We

can see that because this non-negotiable amount of

time (here, 30 seconds) must be provided to serve the

pedestrian calls, barrier times that previously had ac-

ceptable wait times for left-turning trafﬁc are now un-

acceptable. Hence, these barrier times can no longer

be viable when considering pedestrian calls. Thus,

the inclusion of pedestrian calls is vital during simu-

lation. We present the simulation results for the four

critical trafﬁc times we identiﬁed. The ”Major vs.

Minor Wait Times wrt. Barrier Time” plots show

the trade-off for (75th percentile) wait times of ma-

jor and minor streets (Figures 11, 12). The ”Major

and Minor Throughputs wrt. Barrier Time” plot on

the right shows the change in throughput when the

barrier time is changed. Thus, it is possible to un-

derstand the trade-off between the Major and Minor

phase trafﬁc ﬂows in terms of wait times and through-

puts (Figures 11, 12). Ideally, we want to lower

signiﬁcant wait times while not massively increasing

minor phase wait times and while not lowering the

throughput.

Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye Video

Figure 8: Given the ring-and-barrier scheme at this intersection, left turns are affected by the inclusion of pedestrian calls.

The ring structure of the signal timing plan is shown at the bottom of the ﬁgure. The pedestrians cross parallel to the moving

trafﬁc, i.e., left-turning trafﬁc does not interact with pedestrians for safety reasons. Hence, the left-turning phases, i.e., 1/5

and 3/7, get less when pedestrian calls are made since a non-negotiable pedestrian crossing time (here, 30 seconds) must be

given. Further, barrier time restricts the total time the half-rings (i.e., major(1/5 and 2/6) and minor(3/7 and 4/8) halves) get.

Taking an extreme barrier time of 40 seconds means 2/6 must get at least 30 seconds (when pedestrian calls are considered),

leaving just 10 seconds for 1/5. Hence, there is a very high wait time of 642 seconds for left-turning trafﬁc. Had pedestrian

calls not been included in the simulation, the same situation would have given us a wait time of 130 seconds. Hence, it is

important to include the impact of pedestrian calls while analyzing the trade-off between major and minor street wait times.

Figure 9: Replication of intersection geometry and ﬂows in

SUMO simulator. The intersection base map was manually

made. The trafﬁc information was automatically extracted

and fed to SUMO.

6 CONCLUSION AND FUTURE

WORK

Fisheye data provides additional information about

the intersection dynamics, such as accurate turn-

movement counts and ﬂows, vehicle lengths, and

pedestrian information. Unlike regular non-distorted

cameras, of which several are required to capture the

intersection view, a single ﬁsheye lens camera can be

used instead. But there are signiﬁcant issues due to

the ﬁsheye distortion. However, these can be over-

come using the latest computer vision and smoothing

techniques to yield accurate trajectories, speeds, and

Figure 10: Overall trafﬁc throughput shows four high-

volume times for analysis: weekday AM peak, weekday

midday peak, weekday PM peak, and weekend PM peak.

Efforts can be focused on managing the trafﬁc at these

times.

acceleration proﬁles of vehicles and pedestrians.

This data fusion with widely available controller

log data is valuable to trafﬁc engineers and city au-

thorities, as it provides new insights into trafﬁc (ve-

hicular and pedestrian) behavior and safety. These

two data sources augment each other in various ways

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

(a) Simulations results for Weekday AM Peak. In this scenario, we can see that major phase wait times could be decreased by

lowering the barrier time to 70 or 80 seconds, with a relatively small increase in minor phase wait times. The throughput would

remain unchanged. Beyond 70 seconds, there would be a massive increase in minor phase wait times, with no appreciable

improvement in major wait times.

(b) Simulations results for Weekday Midday Peak. In this scenario, we can see that major phase wait times could be decreased

by lowering the barrier time to 70 or 80 seconds, with a relatively small increase in minor phase wait times. The throughput

would remain unchanged. Beyond 70 seconds, there would be a massive increase in minor phase wait times, with no appre-

ciable improvement in major phase wait times.

slightly by lowering the barrier time to 60 or 70 seconds, but with a more signiﬁcant increase in minor phase wait times. The

throughput would remain unchanged. Beyond 60 seconds, there would be a massive increase in minor phase wait times, with

no appreciable improvement in major phase wait times.

Figure 11: Simulation results for Weekday AM peak, Weekday midday peak, Weekday PM peak.

Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye Video

(a) Simulations results for Weekend PM Peak. In this scenario, we can see that Major phase wait times could be decreased

by lowering the barrier time to 60 or 70 seconds, but with a more signiﬁcant increase in minor phase wait times. The

throughput would remain unchanged. Beyond 60 seconds, there would be a massive increase in minor phase wait times, with

no appreciable improvement in major phase wait times.

Figure 12: Simulation results for Weekend PM peak.

and, when used together, can be helpful for safety

analysis and aid data-driven trafﬁc operations.

In the near future, we intend to build an end-to-end

system. It will be capable of supporting multi-camera

ﬁsheye data fusion for large intersections. We also

plan to fuse other modalities of data, such as trajec-

tory data. The system will fully automate the tasks of

cleaning, processing, and visualizing the intersection

dynamics and can be easily deployed across several

intersections with minimal manual effort.

ACKNOWLEDGEMENTS

The work was partly supported by NSF CNS 1922782

and by the Florida Department of Transportation

(FDOT) District 5. The opinions, ﬁndings, and con-

clusions expressed in this publication are those of the

authors and not necessarily those of FDOT D5.

REFERENCES

Alam, M., Ferreira, J., and Fonseca, J. (2016). Introduc-

tion to intelligent transportation systems. In Intelligent

Transportation Systems, pages 1–17. Springer.

Banerjee, T., Chen, K., Almaraz, A., Sengupta, R., Kar-

nati, Y., Grame, B., Posadas, E., Poddar, S., Schenck,

R., Dilmore, J., Srinivasan, S., Rangarajan, A., and

Ranka, S. (2022). A modern intersection data ana-

lytics system for pedestrian and vehicular safety. In

2022 IEEE 25th International Conference on Intel-

ligent Transportation Systems (ITSC), pages 3117–

3124.

Bochkovskiy, A., Wang, C.-Y., and Liao, H.-Y. M. (2020).

Yolov4: Optimal speed and accuracy of object detec-

tion. ArXiv, abs/2004.10934.

Bookstein, F. (1989). Principal warps: thin-plate splines

and the decomposition of deformations. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

11(6):567–585.

Borgi, T., Zoghlami, N., and Abed, M. (2017). Big data for

transport and logistics: A review. pages 44–49.

Buch, N., Velastin, S. A., and Orwell, J. (2011). A review of

computer vision techniques for the analysis of urban

trafﬁc. IEEE Transactions on intelligent transporta-

tion systems, 12(3):920–939.

Carson, J. L. et al. (2010). Best practices in trafﬁc incident

management. Technical report, Federal Highway Ad-

ministration, United States Department of Transporta-

tion.

Chen, K., Banerjee, T., Huang, X., Ding, Z., Varanasi,

V. S., Rangarajan, A., and Ranka, S. (2021). A web-

based visual analytics system for trafﬁc intersection

datasets. International Journal of Big Data Intelli-

gence, 8(1):76–88.

Chui, H. and Rangarajan, A. (2000). A new algorithm for

non-rigid point matching. IEEE Computer Society

Conference on Computer Vision and Pattern Recog-

nition, 2:44–51 vol.2.

DLR (2020). Theory/Trafﬁc Simulations, https://sumo.dlr

.de/docs/Theory/Trafﬁc\

Simulations.html.

Gordon, R. (2016). Intelligent transportation systems.

Cham: Springer.

Huang, X., He, P., Rangarajan, A., and Ranka, S. (2020). In-

telligent intersection: Two-stream convolutional net-

works for real-time near-accident detection in trafﬁc

video. ACM Transactions on Spatial Algorithms and

Systems (TSAS), 6(2):1–28.

Levy, J. I., Buonocore, J. J., and Stackelberg, K. V. (2010).

Evaluation of the public health impacts of trafﬁc con-

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

gestion: a health risk assessment. Environmental

health, 9(1):65.

Lopez, P. A., Behrisch, M., Bieker-Walz, L., Erdmann, J.,

Flotterod, Y.-P., Hilbrich, R., Lucken, L., Rummel,

J., Wagner, P., and Wiessner, E. (2018). Microscopic

Trafﬁc Simulation using SUMO. In IEEE Intelligent

Transportation Systems Conference (ITSC). IEEE.

Lownes, N. E. and Machemehl, R. B. (2006). Vissim:

a multi-parameter sensitivity analysis. In Proceed-

ings of the 2006 Winter Simulation Conference, pages

1406–1413. IEEE.

Mishra, A., Chen, K., Poddar, S., Posadas, E., Rangara-

jan, A., and Ranka, S. (2022). Using video analytics

to improve trafﬁc intersection safety and performance.

Vehicles, 4(4):1288–1313.

Mondal, A., Dutta, A., Dey, N., Sen, S., and Fuqian, S.

(2019). Visual trafﬁc surveillance: A concise sur-

vey. 2019 the 4th International Conference on In-

formation Technology and Intelligent Transportation

Systems, Frontiers in Artiﬁcial Intelligence and Appli-

cations.

Mullakkal Babu, F. A., Wang, M., Farah, H., Arem, B., and

Happee, R. (2017). Comparative assessment of safety

indicators for vehicle trajectories on highways. Trans-

portation Research Record Journal of the Transporta-

tion Research Board, 2659.

OpenStreetMap contributors (2017). Website data . https:

//www.openstreetmap.org.

Rao, A. M. and Rao, K. R. (2012). Measuring urban trafﬁc

congestion-a review. International Journal for Trafﬁc

& Transport Engineering, 2(4).

Santhosh, K. K., Dogra, D. P., and Roy, P. P. (2020).

Anomaly detection in road trafﬁc using visual surveil-

lance: a survey. ACM Computing Surveys (CSUR),

53(6):1–26.

US DOT FHWA (2020). Types of Trafﬁc Analysis Tools.

https:// ops.fhwa.dot.gov/trafﬁcanalysistools/ type\ t

ools.htm.

Weisbrod, G., Vary, D., and Treyz, G. (2003). Measuring

economic costs of urban trafﬁc congestion to business.

Transportation research record, 1839(1):98–106.

Wojke, N., Bewley, A., and Paulus, D. (2017). Simple on-

line and realtime tracking with a deep association met-

ric. In 2017 IEEE International Conference on Image

Processing (ICIP), pages 3645–3649.

Yeh, I.-C., Chung, S., Hsu, Y.-P., and Huang, C.-I. (2020).

Event-based surveillance video retrieval through tra-

jectory analysis using ﬁsheye camera. In 2020 IEEE

International Conference on Consumer Electronics-

Taiwan (ICCE-Taiwan), pages 1–2. IEEE.

Zhang, K. and Batterman, S. (2013). Air pollution and

health risks due to vehicle trafﬁc. Science of the to-

tal Environment, 450:307–316.

Zhao, Z., Zhao, Z., Wang, S., Watta, P., and Murphey, Y. L.

(2021). Pedestrian re-identiﬁcation using a surround-

view ﬁsheye camera system. In 2021 International

Joint Conference on Neural Networks (IJCNN), pages

1–8. IEEE.

Towards Effective Trafﬁc Signal Safety and Optimization Using Fisheye Video