
Research Article
Red-Light Running Violation Detection of Vehicles in Video Using Deep Learning Methods
@INPROCEEDINGS{10.1007/978-3-031-08878-0_15, author={Nam Nguyen Van and Hanh Le Thi and Minh Phan Nhat and Long Lai Ngoc Thang}, title={Red-Light Running Violation Detection of Vehicles in Video Using Deep Learning Methods}, proceedings={Industrial Networks and Intelligent Systems. 8th EAI International Conference, INISCOM 2022, Virtual Event, April 21--22, 2022, Proceedings}, proceedings_a={INISCOM}, year={2022}, month={6}, keywords={Traffic violation detection Vehicle detection and tracking Recurrent neural networks}, doi={10.1007/978-3-031-08878-0_15} }
- Nam Nguyen Van
Hanh Le Thi
Minh Phan Nhat
Long Lai Ngoc Thang
Year: 2022
Red-Light Running Violation Detection of Vehicles in Video Using Deep Learning Methods
INISCOM
Springer
DOI: 10.1007/978-3-031-08878-0_15
Abstract
Recently, Traffic Monitoring Systems (TMS) based on camera are widespread used in many large cities thanks to advances in artificial intelligence especially in deep learning and computer vision. Detection of traffic violation of vehicles is a critical problem for law enforcement in such TMS due to complicated trajectories of different vehicle types in road. Existing methods based on computer vision techniques for detecting, tracking vehicles and then applying violation rules on the perceived path of every vehicles. In this paper, we present a novel approach which is based on the flexible LSTM recurrent neural networks in addition to the traditional fixed rules to detect red-light running violation of vehicles. We also present our improvements on the existing DeepSort tracking algorithm for faster and more accurate ID matching. We evaluate our deep LSTM with attention mechanism on a dataset (Dataset and code are available here:https://github.com/namnv78/RunningRedlight) of 108 traffic videos captured from three road intersections in Vietnam including 628 red-light running violated vehicles. Our method achieved a precision, recall and F1-score of more than 99% which is 3% higher than the traditional rule-based method.