TRACKING THE UNSEEN: AN OCCLUSION-ROBUST FRAMEWORK FOR TARGET TRACKING UNDER FULL AND LONG-TERM OCCLUSION

Multi-object tracking systems often fail when targets disappear behind obstacles, leading to fragmented trajectories and identity loss. This paper addresses the challenge of maintaining continuous target awareness during full and long-term occlusion. The authors propose a three-stage tracking pipeline: (1) YOLOv11n for real-time object detection, (2) an adaptive Kalman Filter for motion prediction during visibility gaps, and (3) the Occlusion-Aware Mask Network (OAMN) for appearance-based re-identification (Re-ID). The framework specifically uses an IoU-gate bypass mechanism that triggers appearance-only matching when a target remains occluded for more than 30 frames, allowing the system to recover the correct identity upon reappearance.

The authors conducted a systematic evaluation of six different Re-ID architectures within their pipeline. OAMN was selected as the optimal module because its two-branch architecture—which learns to mask out occluded regions at the feature level—prevents background noise from corrupting the target's appearance embedding. When benchmarked against the state-of-the-art OccluTrack system on the public OVIS dataset, the proposed framework achieved relative improvements of 18.1% in Multiple Object Tracking Accuracy (MOTA) and 25.1% in Identity F1 Score (IDF1), while reducing identity switches by 12.8%. Furthermore, in a custom military dataset simulating battlefield surveillance, the framework demonstrated robust performance with a MOTA of 0.734, confirming its effectiveness for defense applications where targets frequently disappear behind structures.

In high-stakes environments like military surveillance, losing track of a target even briefly can lead to mission failure or incorrect situational awareness. By successfully integrating motion prediction with occlusion-aware appearance modeling, this framework provides a reliable solution for maintaining persistent identity across extended periods of invisibility. The study demonstrates that combining lightweight, efficient detectors with specialized Re-ID architectures is a viable path toward building robust, real-time tracking systems for complex, cluttered environments.