Comparative Evaluation of SORT, DeepSORT, and ByteTrack for Multiple Object Tracking in Highway Videos
Mahmoud Abouelyazid
Purdue University
Keywords: multiple object tracking, SORT, DeepSORT, ByteTrack, highway timelapse, traffic monitoring, autonomous driving, video surveillance
Abstract
Multiple object tracking is a fundamental task in computer vision with significant implications for various applications, including traffic monitoring, autonomous driving, and video surveillance. This study aims to compare the performance of three state-of-the-art tracking algorithms: SORT, DeepSORT, and ByteTrack, in detecting and tracking vehicles and persons in highway timelapse videos. SORT is a simple and efficient tracking framework that combines detection and tracking to estimate object states in real-time. DeepSORT extends SORT by incorporating deep learning techniques to reduce identity switches and enhance tracking accuracy. ByteTrack, in contrast, is a one-shot detection-based approach that integrates object detection and tracking into a single model for improved efficiency. To evaluate the performance of these tracking methods, we employ a set of evaluation metrics, including Multiple Object Tracking Accuracy (MOTA), Multiple Object Tracking Precision (MOTP), ID Switches (IDs), Mostly Tracked (MT), Mostly Lost (ML), False Positives (FP), False Negatives (FN), and Processing Speed. The experiments are conducted using an open access video dataset. The experimental results reveal that ByteTrack consistently outperforms SORT and DeepSORT across most evaluation metrics. ByteTrack achieves a MOTA of 77.3%, MOTP of 82.6%, and the lowest number of ID switches at 558. It also demonstrates the highest percentage of mostly tracked objects (54.7%) and the lowest percentage of mostly lost objects (14.9%). Moreover, ByteTrack maintains a high processing speed of 171 FPS, surpassing both SORT and DeepSORT in terms of computational efficiency. This research shows the better performance of ByteTrack in accurately and efficiently tracking multiple objects, vehicles and persons, in highway timelapse videos. The findings of this research have implications for the development of robust and real-time tracking systems for various intelligent transportation and surveillance applications. Future research directions include further optimization of the ByteTrack algorithm and its adaptation to real-world scenarios.
