Performance Analysis of YOLO-Deep SORT on Thermal Video-Based Online Multi-Objet Tracking

Recently, thermal cameras have been used in various fields, including surveillance systems and advanced driver assistance systems (ADAS), as they perform better in low light than visible-light cameras. Some challenges in the surveillance system or ADAS field related to thermal cameras are occlusion and thermal crossover between objects with similar appearances during object detection or object tracking tasks, which can lead to misdetection, false positives, and lost tracking. In this paper, performance analysis of you-only-look-once (YOLO) combined with deep online real-time tracking (DeepSORT) on thermal video-based online multi-object tracking (MOT) in occlusion and thermal crossover scene is presented. YOLO, as one of state-of-the-art method for detection task, is used for detection system. Then, the detected object from YOLO is tracked using DeepSORT. The results demonstrate that the online MOT of sequential thermal images using YOLO-DeepSORT achieved a MOTA score of 44.2% and IDF1 of 45.3%. Thus, negative example was added in YOLO training process to reduce false detection, and it gives improvement with MOTA score of 63.8% and IDF1 score of 54.6%.