A Crowdsourcing Approach to Tracker Fusion

There are many tracking methods been proposed, using different features and algorithms, but none of them can track object correctly all the time. In this paper, we explore the idea of combining a crowd of trackers, and propose a crowdsourcing tracking method. We model the problem under the Sequential Monte Carlo framework, where we treat different trackers outputs, the bounding boxes, as weak observations, and use the wisdom-of-the-crowds to simultaneously infer both the hidden ground truth bounding box and the corresponding time-varying confidence for each tracker. We have tested our proposed method on two public surveillance video datasets and two of our own video datasets. The results show that the crowdsourcing tracking method can provide more stable and better performance.