Performance Evaluation of Visual Object Tracking using YOLO Deep SORT with LCF

Abstract

Most of the problems in object tracking are illumination and increased ID switching because of the occlusion and re-entering. To overcome the problems of object tracking, the proposed system uses the You Only Look Once (YOLO) and Deep Simple Online and Real-Time Tracking (Deep SORT) with Low Confidence Track Filtering (LCF) for object tracking. As a single-stage detector, YOLO performs classification and bounding box regression in one step, it makes faster than most convolutional neural networks. To get better performance for tracking and re-identification of the ID switch, Deep SORT with LCF is applied in this system. When filtering the low average confidence by LCF, the false positive tracks can be reduced. In this paper, YOLOv3 and YOLOv4 are comparatively performed for the dedicated visual tracking system. According to the experiment results for long-term occlusions, YOLOv4 and Deep SORT with LCF can identify more objects tracking than YOLOv3 and Deep SORT with LCF.