Improved Algorithm for Road Multi-target Tracking Based on YOLO

Abstract

Based on the single-frame target detection results of the YOLO algorithm, an improved algorithm for multi-target continuous tracking in road scenes is proposed. Firstly, the characteristics of the target eigenvalues of the YOLO algorithm are analyzed, and the judgment basis for using the difference degree to track the target between frames is clarified. Secondly, the defect of the original Euclidean distance as the difference degree calculation is analyzed, and a new difference degree definition is proposed; the value rule of the target feature vector in the road scene is analyzed, and the conversion function is introduced. Next, considering the prior knowledge in the road scene, the difference degree and the intersection ratio and IOU are used to jointly determine and optimize the algorithm process. Finally, according to the experimental comparison, it is confirmed that in the road scene, the new improved algorithm can effectively improve the correct rate of multi-target recognition and reduce the false detection rate and missed detection rate while basically maintaining low time-consuming operation.