Research on Improved DeepSort Pedestrian Detection Algorithm

An algorithm that combines improved YOLOv5 and improved DeepSort to address the problem of pedestrian multi-target tracking in complex scenes. Firstly, we choose the more efficient YOLOv5 as the basic detection network, while embedding the ECANet attention module to enhance feature extraction capability and enhance detection accuracy. Secondly, in the DeepSort tracking algorithm, we introduce a lightweight GhostNetV1 to replace the original re identification network, reducing computational complexity and improving the performance of pedestrian re identification.To optimize the matching process, we use the Hungarian algorithm for optimal matching and use DIOU instead of IOU for secondary matching, in order to more accurately reflect the similarity between the detection box and the prediction box, and improve tracking stability. The experimental results show that the new algorithm performs better in scenarios such as occlusion, dense crowds, and poor lighting, with significantly reduced false positives and missed detections, and significantly enhanced robustness.In addition, due to the use of lightweight networks and efficient matching strategies, the new algorithm reduces computational costs and improves real-time performance while maintaining high performance. This improvement provides strong support for the promotion of pedestrian multi-target tracking in practical applications, and provides new ideas and methods for solving tracking problems in complex scenes.