KFM MOTR

Multi-object tracking (MOT) in complex video sequences presents significant challenges, particularly under conditions of occlusion, object deformation, and intermittent visibility. Existing transformer-based MOT methods don’t perform well in the aforementioned scenarios. This paper introduces an advanced MOT framework that integrates a Kalman filter, a novel memory module, and a state-of-the-art (SOTA) object detector to address these challenges. Our proposed system employs the Kalman filter to predict the trajectories of occluded objects, enhancing tracking continuity even when direct detections fail. The memory module facilitates robust re-identification of objects that undergo deformation or disappear and reappear, leveraging stored feature histories for accurate association. Additionally, the integration of a SOTA object detector optimizes initial detections, setting a strong foundation for tracking accuracy. We present a comprehensive evaluation of our system on standard datasets, including comparisons with recent state-of-the-art methods like MeMOTR and MOTRv2. Our results show a significant improvement, with a 4.86% increase in the Higher Order Tracking Accuracy (HOTA) metric, underscoring the effectiveness of our approach in handling real-world tracking scenarios. The proposed enhancements not only advance the performance of MOT systems but also provide a scalable foundation for future research into more resilient and adaptive tracking technologies.