A Decentralized Collaborative Strategy for PTZ Camera Network Tracking System using Graph Learning: Assessing strategies for information sharing in a PTZ camera network for improving vehicle tracking, via agent-based simulations

In this paper, we propose a collaborative strategy for an autonomous PTZ (Pan Tilt Zoom) camera network, for improved target vehicle tracking, in real-world traffic. In addition, we explore ways to improve the target-tracking system to adapt itself to unexpected changes in traffic patterns. In the exploration phase, the camera nodes utilize the target identification information shared among themselves in the network to learn a graph. The vertices represent the camera nodes, while edges represent links to immediate neighbors, and edge weights represent the distance between the nodes in terms of the time taken by the target vehicles. In the exploitation phase, once a target vehicle is identified, the camera broadcasts the information to the neighbors. In turn, the neighbors consult the graph to position themselves better for capturing footage of the target vehicle. We carried out two agent-based simulation experiments to test the strategy. In the first experiment, we compare the proposed strategy, which uses the learned graph, to a baseline where the cameras operate independently for scanning traffic. In the second experiment, we compare the strategy to an improved adaptive version of itself, in which the system learns online continuously by observing live traffic. The results show that our cooperative camera network outperforms the baseline and the adaptive strategy outperforms the static one.