Data Association Tools for Target Identification in Distributed Multi-target Tracking Systems

Distributed tracking systems have several benefits over centralized setups such as faster processing time and greater robustness to failures. However, the practical deployment of a distributed multi-camera multi-target tracking system poses other important challenges. In this work, we address two of these practical problems. The first one is the spatial and temporal identification of the targets in the network, i.e., the data association problem. To solve it, we propose to build intelligent and adaptive local appearance models of each target that only store the most relevant information. The second problem is the intensive use of bandwidth caused by the periodic communications that each camera requires for the cooperative tracking and the data association of all the targets. In the paper, we manage the bandwidth usage with an event-triggered mechanism that controls how much information is sent. The main novelty of our mechanism is to account for the scene density, coupling it with the data association module and enhancing it. We integrate the new modules into an existing distributed multi-person multi-camera tracking system and demonstrate their benefits on different public benchmarks of increasing difficulty.