Unsupervised path modeling across multiple cameras with disjoint views for foreground object tracking

We present an unsupervised path modeling method to track foreground objects across multiple cameras with disjoint views. To avoid the training process in most approaches, our method imposes the camera topology graph to identify possible paths of foreground objects in the camera network. Then, with the appearance and temporal information, dynamic programming is applied to search an optimal path of a target foreground object with a given length of the path. The experimental results show that the proposed method can successfully track target objects across multiple cameras with disjoint views compared to the state-of-the-art supervised methods.