Abstract: based detection method is proposed. First，a transformer module is introduced into the convolutional module of the YOLOv5s network to capture richer global information. Then, the path aggregation network and loss function of the YOLOv5s network are optimized to improve its detection speed and accuracy，resulting in an enhanced YOLOv5s network. Finally, autonomous driving object images in complex environments are input into the improved YOLOv5s network for detection，and simulation experiments are conducted for validation. The results demonstrate that the improved YOLOv5s network achieves mean average precision, recall， parameter count, computational complexity，and detection speed of 97.62%，94.71%，18.15M，39.57G，and 35.35 Hz，respectively， for autonomous driving target detection in complex scenarios，exhibiting superior performance. Consequently, this approach enhances the accuracy of autonomous driving target detection in complex environments，providing foundational technical support for ensuring the safety of autonomous driving operations.