Decentralized Visual Tracking of Multiple Targets
Tracking multiple targets in video is critical for video monitoring and surveillance. Tracker coalescence, meaning that some targets have multiple tracks while some targets have none (or are lost), is a challenging and common problem in practice. Existing solutions that are based on joint data association are centralized schemes and are in general confronted by their combinatorial complexities.
We give a novel decentralized solution that has a linear complexity and can be easily parallelized so that the accuracy and the capacity of the multiple target tracker can be significantly enhanced. Our decentralized/collaborative solution employs a set of collaborative autonomous trackers in a dynamic ad hoc random field where exclusive motion priors are designed and embedded. This new model enables the competition of these individual trackers that converge at a mean field fixed-point. Our experiments indicate that even simple competition models can lead to significant enhancement in performance. Recently, we have extended this new method to handle a variable number of targets. The new approach that largely overcomes coalescence with linear complexity significantly advances the state-of-the-art of multiple target tracking, and may greatly impact practical problems.
|
|
|
|
|
|
|
|
|
|