The pedestrian detection algorithm based on an improved YOLOv8s

To address the issues of low accuracy and large parameter size in the original YOLOv8 model for pedestrian detection, we propose an improved object detection algorithm named DB-YOLOv8s, which is based on the YOLOv8s model. First, we enhance the Neck structure of YOLOv8s using BiFPN, improving the network's multi-scale feature fusion capability. We also introduce DCNv2 convolution to replace the standard convolution in the Backbone. Additionally, we refine the C2f structure with DCNv2, proposing a C2f_DCNv2 structure to enhance the network's ability to learn and represent human targets. Finally, we integrate Dyhead to boost the network's capability in detecting small objects. Our model was trained on the CrowdHuman dataset and tested on WiderPerson, showing significant improvements compared to the original YOLOv8 model.