Online Multi-Object Thermal Tracking and Reidentification using YOLO and DeepSORT in Low Light Environment

Multi-object tracking and reidentification has been used in various domains such as surveillance, human-machine interface, and vehicle navigation. However, existing methods primarily rely on deep learning and tracking-by-detection (TBD) techniques using visible light cameras, which face challenges in low-light conditions. To address this, we propose a novel multi-object tracking approach utilizing thermal cameras to overcome visibility issues in such environments. Our method integrates the YOLOv7 for precise image detection and the DeepSORT algorithm for efficient tracking amidst occlusion and object similarities. Through rigorous experimentation, we achieved notable results, obtaining an IDF1 score of 0.911, F1 Score of 0.878, MOTA of 0.767, and MOTP of 0.161, underscoring the efficacy of our approach in multi-object tracking under challenging lighting conditions. Given the scarcity of publicly available thermal datasets, we curated a thermal reidentification dataset comprising 10,000 thermal images captured in diverse low-light settings. Our dataset includes various poses and perspectives to enhance model performance. We structure our study to present related works, detail our proposed methodology, and analyze experimental results, highlighting the efficacy of our approach in multi-object tracking under low-light conditions.