MULTI-OBJECT TRACKING METHOD WITH ADAPTIVE OCCLUSION PROCESSING FOR VIDEO SURVEILLANCE SYSTEMS

Abstract

This paper presents an improved multi-object tracking (MOT) method based on the ByteTrack framework, with a primary focus on enhancing robustness to occlusions in complex and dynamic video scenes. Occlusions, camera motion, and dense object interactions remain critical challenges for real-time tracking systems, often leading to identity switches and track fragmentation. To address these issues, the proposed approach introduces two key improvements: the integration of the Distance Intersection over Union (DIoU) metric for more precise detection-to-track association, and a track retention mechanism that allows temporarily lost objects to be preserved during short-term or partial occlusions.

The use of DIoU enables more reliable association decisions by jointly considering both spatial overlap and the distance between object centers, which is especially beneficial in crowded environments where bounding boxes frequently intersect or overlap. In addition, the proposed track retention strategy maintains inactive tracks for a limited period, allowing the system to recover object identities once they reappear, thereby reducing premature track termination.

Experimental evaluation was conducted on the MOT17 validation dataset using a lightweight YOLOX-Tiny detector to ensure real-time applicability on resource-constrained platforms. The results demonstrate that the proposed enhancements lead to consistent performance gains compared to the baseline ByteTrack algorithm. Specifically, the method achieved an improvement in MOTA of +0.5% (from 77.5% to 78.0%), a +1.9% increase in IDF1, and a 9% reduction in identity switches, while maintaining real-time processing with only a marginal decrease in speed (from 29.6 to 29.1 FPS).

Additional scene-based analyses confirmed improved robustness in scenarios involving dense crowds, long-term occlusions, and camera motion. Overall, the results indicate that efficient occlusion handling can significantly enhance tracking reliability and identity consistency without sacrificing computational efficiency. The proposed method offers a well-balanced trade-off between accuracy and performance, making it suitable for practical deployment in video surveillance systems, intelligent transportation monitoring, and mobile robotic applications operating under real-time constraints.