Dynamic task decomposition for decentralized object tracking in complex scenes

摘要: Decentralized tracking in multicamera environment is formalized as assignment problem.Assignment decomposes task into subtasks under occlusion constraints.Assignment found by minimizing a parameter-free objective function F.F derived via information theoretic measure of predicted uncertainty assignment.Min-cost flow solver exploits graph encoding structure F and constraints. The employment visual sensor networks for video surveillance has brought many challenges advantages. While the integration multiple cameras network potential advantage fusing complementary observations from sensors enlarging coverage, it also increases complexity tasks poses to system scalability. For real time performance, key approach tackling these mapping global onto distributed sensing processing infrastructure. In this paper, we present an efficient scalable multi-camera multi-people with three-layer architecture, which formulate overall (i.e., all people using available cameras) vision based state estimation problem aim maximize utility sharing resources. By exploiting geometric relations between geometry people's positions, our method able dynamically adaptively partition number nearly independent aid reasoning, each tracks subset (or agencies). hereby reduces dramatically helps boost parallelization system's throughput reliability while accounting intrinsic induced, e.g., clutter occlusions. We demonstrate efficiency decentralized tracker on challenging indoor outdoor sequences.