Re-ID Free Multi Camera Tracking and Real-Time ID Fusion for ROS2-Based Autonomous Parking Systems

Vehicle tracking using fixed cameras is a key technology in smart parking and urban mobility systems. However, in environments where vehicles share near-identical appearance and lack explicit identifiers such as license plates, existing visual-based Re-ID methods often fail, leading to frequent identity switching. To overcome this, we propose a Re-ID-free multi-camera tracking framework that leverages geometric projection, motion priors, and probabilistic data association. Our method projects local tracklets into a unified bird's eye view (BEV) coordinate system using homography transformations. Kalman filters with motion priors are used to model trajectory continuity, while Mahalanobis distance-based gating filters improbable associations. Joint Probabilistic Data Association (JPDA) resolves ambiguous matches under occlusion and overlapping fields of view. The system is implemented in real-time on ROS 2 and evaluated in a scaled indoor parking testbed with visually identical vehicles. Experimental results show significant reductions in ID switching and improvements in tracking consistency. This work presents a practical and scalable tracking solution for dense and constrained environments where appearance-based methods struggle.