Department of Computer Engineering

This study aims to construct a Real-Time Multi-Camera Multi-Person Tracking (MC-MOT) system which jointly optimizes local (single-camera) and global (multi-camera) feature distances. While most existing approaches follow a two-stage track-then-associate scheme, this work focuses on a joint approach. Our method also operates in real-time in contrast to the more common offline or windowed joint tracking algorithms which operate on future information. In summary, this study contributes: (i) A joint MCMOT formulation where the optimization objective solves both local and global tracking at teach step, (ii) a realization of the method in the form of an algorithm capable of producing real-time track IDs, and (iii) a new MCMOT evaluation metric we call Global IDF1 which acts as a multi-camera extension of the IDF1 metric, emphasizing continuous traceability of a target across a multi-camera network. We further propose a Multi-View Fusion (MVF) network to extract descriptive feature vectors for multi-camera detection groups. We report results comparable to offline state-of-the-art methods while remaining real-time and retaining simplicity.