Enhanced Deep SORT with Class Filtering for ID Switches Minimization in Visual Object Tracking

A new approach for tracking the visual objects is proposed to enhance Deep SORT with class filtering in this system. Currently, visual object tracking systems are most popular in real life. But there are still challenging tasks such as illumination change, occlusion and reappearing on camera's view, and the tracked objects’ ID switching. Moreover, in the tracking process, false positive tracks are changed the tracking object's ID, and different detection class and tracking class are also problems for ID switching cases. The main emphasis of this paper is to minimize IDs’ changes of objects. In this paper, low-confidence, and proposed class filters are implemented and tested on own dataset to enhance Deep SORT. In the experimental results, the changes in tracking specified objects’ ID of the two filtering can minimize by 10% and 67% respectively rather than that of the Deep SORT. Therefore, we found that a new approach can significantly minimize the IDs’ changes of objects.