A computer vision‐based deep learning model to detect wrong‐way driving using pan–tilt–zoom traffic cameras

Abstract

Hundreds of fatal accidents occur each year due to wrong-way driving (WWD). Although several methods have been developed to detect WWD using existing closed-circuit television (CCTV) data, they all require manual recalibration whenever a camera rotates, and are thus not scalable across statewide CCTV networks. This paper, therefore, proposes an end-to-end deep-learning-based model that considers camera orientation as a variable, detecting camera rotation automatically and learning new decision criteria accordingly using a neural network model. We show that our proposed solution can detect WWD with a precision of 0.99 and a recall of 0.97. Due to its cheap computational cost and high error tolerance, our solution is easily scalable for statewide surveillance on a real-time basis to help decision-makers reduce fatalities due to WWD. INTRODUCTIONBased on the National Highway Traffic Safety Administration's Fatality Analysis Reporting System, the annual average of 269 fatal crashes from 2004 through 2011 due to wrong-way driving (WWD) resulted in more than 350 people dying each year (Pour-Rouholamin et al., 2015).According to federal and state crash data, 20% to 25% of crashes from WWD are fatal. This percentage is significant, compared to the 0.5% fatality rate, for vehicle crashes overall (MH Corbin, 2020).Several studies (such as Baratian-Ghorghi et al., 2014;Leduc, 2008;Zhou et al., 2012) have investigated different factors in WWD accidents and have shown that in 58% of the crashes, drivers were under the influence of alcohol.To prevent drivers from entering the wrong side of the road, several passive mechanisms have been implemented, such as wrong-way signs or raised pavement markers on roads (Baratian-Ghorghi & Zhou, 2017;Cooner et al., 2003) and in-pavement warning lights. In parallel with deploying passive mechanisms to minimize WWD incidents, many

A hybrid virtual–real traffic simulation approach to reproducing the spatiotemporal distribution of bridge loads

“…The combination of YOLO and SORT algorithms for object detection and trajectory tracking, respectively, is widely used in computer vision (Haghighat & Sharma, 2023; Veeramani et al., 2018). In this study, YOLOv5 and DeepSort were used for multi‐object detection and object trajectory tracking, respectively.…”

Enhanced Real-Time Computer Vision and Intelligent Decision-Making for Autonomous Vehicle Applications

“…Computer vision without an efficient and robust deep learning framework would yield irregularities and unusable outputs. For this, frame rate as well as mAP score provides the most fundamental basis for any detection application whether vehicles, animals, transport entities or even the traffic signs [41][42]. Additionally, from the safety perspective also, the detection of correlated entities is essential which could only be achieved from high frame rate and high mAP scores [43].…”

Object Detection in Traffic Videos: A Survey

“…Although this SDK is closed-source, developers have the option to add implementations in C, C++, or Python, integrate custom libraries, and deploy models for inference in native frameworks such as TensorFlow and PyTorch. This framework has been used for traffic surveillance applications in a number of recent studies [167], [236]- [238].…”