Optimization of Traffic Participant Detection Algorithm in Intelligent Roadside Scene

In the intelligent transportation roadside object detection scene, due to the limited computing power of edge computing equipment, it is difficult to deploy large and complex object detection models, at the same time, most of detection models can not give consideration to precision and real-time. The complex road environment requires the model to have higher detection precision and faster detection speed. To solve this problem, an improved lightweight object detection model based on YOLOv4 is proposed(STDC_YOLO). The proposed model uses short-time dense network structure (STDC) to replace cross stage partial model structure (CSP), employs multi-scale feature fusion module which combines low-level feature and high-level feature to enrich feature information. In addition, the proposed STDC_YOLO employs re-parameterized VGG block (RepVGG) to improved parameter efficiency, thereby improving the inference speed. In order to improve the positioning precision of the predication bounding box, the SIOU loss function is employed to accelerate the convergence of the model and improve the positioning precision of the predication bounding box. Experiments show that STDC_YOLO network proposed in this paper has better metrics on the roadside traffic participant dataset. Compared with YOLOv7-tiny object detection algorithm, the precision is increased by 2%, and mean average precision(mAP) is increased by 8%. Compared with the original YOLOv4, the optimized model has lower running memory and higher inference efficiency. The evaluation on the MS COCO dataset shows that the model still has good robustness in different detection tasks.