Multi Pedestrian Tracking Algorithm Based on Improved YOLOv5 and DeepSORT

A pedestrian detection and tracking algorithm based on improved YOLOv5s and DeepSORT is proposed to address the issues of pedestrian identity jumps and tracking failures caused by mutual occlusion, similar appearance features, target deformation, and motion blur on traffic roads. In the detection stage, the fusion of attention module ECA and YOLOv5 neck network enhances the extraction of pedestrian features. Replacing the CIoU bounding box loss function with the EIoU bounding box loss function accelerates bounding box regression while improving accurate localization. During the tracking phase, DeepSORT is improved by utilizing an extended Kalman filter to predict pedestrian positions in nonlinear environments. The Hungarian algorithm is used to match and detect trajectories, optimizing the problem of frequent identity switching in complex environments. Finally, the improved YOLOv5 and DeepSORT algorithms were connected to detect and track the MOT-16 dataset. The experimental results show that the improved YOLOv5 algorithm has an average accuracy increase of $3.5 \%$ compared to the original algorithm. When combined with DeepSORT tracking, the average tracking accuracy is $60.5 \%$, which is $3.1 \%$ higher than the original algorithm.The number of pedestrian identity switches decreased by 46 times, which is $5.2 \%$ lower than the original algorithm.