Dense Small Object Tracking via DeepSort Integrated with YOLOv.X-based Detection

In order to address the problem of incomplete feature extraction, missed tracking, and incorrect tracking in dense small object tracking tasks, This paper proposes a DeepSort method based on YOLOv.X for input detection. This method uses YOLOv5 as the detection input for DeepSort to obtain more complete detection box information for dense small objects; Then, the feature extraction and fusion parts in YOLOv5 are reused to replace the appearance feature extraction network in DeepSort, improving the algorithm's performance in extracting features from dense small objects; On this basis, the CIoU algorithm is used as a measure of the correlation between unmatched detection boxes and unmatched prediction boxes during the tracking process to avoid unnecessary trajectory loss. The experimental results show that compared with the original DeepSort tracking algorithm, the improved algorithm improved the MOTA index by 3.6%, the MOTP index by 2.9%, and reduced the error ID conversion index by 269 times.