Multi-object Tracking based on improved YOLO
Zhanbo Li, Yuhong Mu, Tingting Li, Jie Su · 2023
To improve tracking by detection algorithm missed detection and IDswitch problems caused by the occlusion of the target and the change of light and darkness. We propose a multi-target tracking algorithm based on improved YOLOv5 and DeepSORT. The SE and SimAM attention are adapted to YOLO v5 to improve the detector's precision, and the FocalEIOU loss function replaces the original CIOU to further accelerate the regression rate of the bounding box and improve the localization accuracy of the detector for the target. SAConv is added to the YOLOv5 backbone network and combined with the C3 layer, C3-SAC is proposed to address the unsuccessful identification problem caused by target occlusion, The feature extraction model of DeepSORT is adjusted and DIOU is used for matching. The improved algorithm is evaluated for detection and tracking on datasets VOC07+12, MOT16 and MOT20 respectively, and the mAP of VOC grows by 8.57%, MOTA of MOT16 and MOT20 grow by 3.88% and 6.45% respectively, and FPS meets with real-time tracking requirements. The result proves the effectiveness of the improved algorithm.