Research on Detection Method of Traffic Anomaly Based on Improved YOLOv3

摘要

Aiming at the characteristics of the detection of abnormal traffic incidents with large target scale changes, high requirements for location positioning accuracy, and high real-time requirements, a method for detecting abnormal traffic incidents is proposed. In terms of vehicle detection, K-means dimensional clustering is performed on the a priori box size on the multi-scale prediction branch feature map, which enhances the scale adaptability. The loss function of boundary box is improved to improve the locating ability. The channel attention and spatial attention structure (CBAM) is embedded in front of each shortcut layer of the YOLOv3 basic network Darknet53 to expand the network’s perception of the target feature area. In terms of vehicle tracking, the pedestrian re-identification network in Deep SORT is replaced with a vehicle re-identification network, which effectively reduces the ID switch of the target. The experimental results show that the proposed abnormal event identification rules can accurately detect pedestrians on elevated roads, abnormal parking and reversing events in real time.