Aiming at the problems of low target detection accuracy, poor real-time multi-target tracking and difficulty in recognizing small targets in complex traffic scenes, this paper proposes a real-time target detection algorithm based on improved YOLOv5s. A directed graph scene model containing environment features and object features is constructed, and the Marginalized Kernel algorithm is used to enhance the dynamic environment sensing ability. Improve the model architecture of YOLOv5s and optimize the feature extraction with the help of MHSARM. Enhance the spatial localization by combining CoordConv, and realize the joint learning of target detection and epigenetic features based on JDE paradigm. Experimental results show that on the TT100K dataset, the model in this paper outperforms all comparative models, with a 22.92% improvement in mAP@0.5 compared to the YOLOv5s baseline model, achieving an accuracy of 86.17%, and also demonstrating the best detection performance on the BDD100K dataset. The improved model performed best in terms of AP@0.5 accuracy in ablation experiments, achieving a mAP value of 80.24% in validation across six types of real traffic scenarios.
