Research on Small Target Road Detection Method for Autonomous Driving Based on YOLOv5

With the rapid development of artificial intelligence and computer vision technology, the importance of road object detection in autonomous driving and intelligent transportation systems has become increasingly prominent. However, the existing YOLOv5 algorithm has the problem of insufficient accuracy in small object detection, which is difficult to meet the real-time and accuracy requirements in complex road scenes. In this paper, an improved YOLOv5 network structure is proposed, which significantly improves the efficiency of feature extraction by optimizing the Spatial Pyramid Pooling module (SPP) in the backbone network to a lightweight Serial Spatial Pyramid Pooling Module (SPPF). The Adaptive Feature fusion mechanism (ASFF) is introduced into the neck network to effectively enhance the weight of shallow detail features. At the same time, the neck network and the head network are merged to construct an end-to-end detection framework, which further improves the detection performance of small targets. The TT100K and VOC datasets were used for verification. The Mean Average Precision (mAP) of the improved algorithm in small object detection reaches $\mathbf{9 0 \%}$, which is $\mathbf{1 5 \% h i g h e r}$ than that of the original YOLOv5. The research results show that the proposed improved method has higher detection accuracy in complex road scenes, and provides effective technical support for automatic driving environment perception.