Research on the improved YOLOv5 object detection algorithm for autonomous driving scenarios

In autonomous driving, object detection must run efficiently on embedded devices, posing challenges in both model lightweighting and accuracy. Object information is often small and can be lost during detection, making traditional models inadequate. To address this, we propose an improved YOLOv5-based method. The backbone uses ShuffleNetV2 for lightweight design, and Shifted Patch Downsampling convolution is introduced in the neck to prevent fine-grained information loss. Extensive experiments on the Argoverse traffic dataset show that the proposed method outperforms advanced models like YOLOv5 in both detection accuracy and efficiency, providing an effective solution for object detection in autonomous driving.