Data association for multi-object Tracking-by-Detection in multi-camera networks

Multi-object tracking is still a challenging task in computer vision. We propose a robust approach to realize multi-object tracking using multi-camera networks. Detection algorithms are utilized to detect object regions with confidence scores for initialization of individual particle filters. Since data association is the key issue in Tracking-by-Detection mechanism, we present an efficient greedy matching algorithm considering multiple judgments based on likelihood functions. Furthermore, tracking in single cameras is realized by a greedy matching method. Afterwards, 3D geometry positions are obtained from the triangulation relationship between cameras. Corresponding objects are tracked in multiple cameras to take the advantages of multi-camera based tracking. Our algorithm performs online and does not need any information about the scene, no restrictions of enter-and-exit zones, no assumption of areas where objects are moving on and can be extended to any class of object tracking. Experimental results show the benefits of using multiple cameras by the higher accuracy and precision rates.