Multi-camera multi-object tracking (MC-MOT) has become pivotal in various real-world applications within computer vision. Despite extensive research, solving the data association problem remains one of the most formidable challenges in MC-MOT pipeline. This challenge is compounded by factors such as varying illumination, diverse walking patterns, and trajectory occlusions. In recent years, graph neural networks (GNNs) have emerged as promising tools for enhancing data association. However, prevalent graph-based MC-MOT methods often rely on computationally inefficient min-cost flow approaches for cross-camera association, with static graph structures that lack adaptability to new detections. Moreover, these methods typically process cameras in pairs, leading to localized solutions rather than a holistic global approach. To address these limitations, we propose a two-stage lightweight crosscamera tracker designed to achieve a global solution efficiently. Our approach prioritizes the quality of local tracklets, enhancing them through supervised learning on multi-source datasets using the DeepSort model. For multi-camera association, we leverage the dynamic connectivity of Message Passing Graph Neural Networks (MPGNNs) to jointly learn features and similarities previously untapped in this domain. Our proposed model significantly improves detection accuracy and feature extraction, outperforming current MC-MOT algorithms on cross-camera benchmarks. This advancement marks a notable step forward in the field, offering more precise tracking capabilities and demonstrating the potential of integrating state-of-theart techniques for enhanced performance in complex tracking scenarios.
