KAD-SORT: Multi-target tracking algorithm for autonomous driving scenarios based on BoT-SORT

The text discusses the challenges faced by multi-object tracking algorithms in autonomous driving scenarios, such as targets being too small to detect due to being far from the camera, targets being occluded due to dense arrangements or complex backgrounds, and missed detections in low light conditions. An improved multi-object tracking algorithm called KAD-SORT is presented, building upon the BoT-SORT algorithm. First, the YOLOX detection head in the BoT-SORT network is replaced with YOLOv8, and the YOLOv8 network is enhanced by adding a CA attention module and a small-object detection layer to enhance detection precision. Second, acceleration parameters are introduced into the Kalman filtering block of the BoT-SORT network to enhance prediction precision for targets with non-uniform motion trends. Finally, GIoU is introduced as a method for measuring the distance between object boxes in adjacent frames, replacing the IoU association matching mechanism to improve the accuracy of matching predicted boxes with detected boxes. The experimental findings indicate that compared to the BoT-SORT, the KAD-SORT tracking algorithm improves MOTA by 2.6% and IDF1 by 1.7%.