Research on Traffic Accident Detection Based on Vehicle Perspective
Zhuofei Xia, Jiayuan Gong, Yue Long, Wenbo Ren, Jingnan Wang, Haitao Lan · 2022
With more people owning cars, traffic accidents happen more frequently and result in many more fatalities each year. It is challenging for current roadside traffic incident detection equipment to have many blind spots, especially after an accident, which results in secondary injuries from traffic events not being discovered in time. As a result, a technique for detecting traffic incidents is presented for use on the vehicle side and is based on video captured by the car camera. The backbone of the proposed enhanced YOLOv5l model for the detection model is the EfficientNet structure with the SE layer eliminated, which drastically decreases the number of model parameters and computing work. To increase model detection accuracy, the fused attention technique is implemented at the prediction side. When compared to the YOLOv5s model, the upgraded model's mAP improves by 13% to 87.7%; when compared to the YOLOv5l model, our model's detection speed reduces by 7ms to 22.7ms, allowing it to be placed at the vehicle side for real-time detection of road traffic incidents.