Pedestrain Detection Based on Improved CSP Network

This paper presents a method based on improved Cross Stage Partial (CSP) network. Firstly, the original backbone network Resnet50 of CSP is replaced by Resnet101, which better extracts features and solve the problem of pedestrian detection with occlusion. Secondly, a channel-based and compressed attention mechanism is introduced to obtain higher training speed and detection speed. Finally, the non-maximum suppression algorithm is used to form the candidate frames and retrain difficult samples. The experimental results show that the proposed approach achieves better results on the Cityperson dataset and improves the performance of pedestrian detection in complex scenes.