Multi-human tracking from sparse detection responses

Abstract

In this study, the authors focus on the performance improvements of multi-human tracking from sparse detection responses. Many previous detection-based data association tracking methods used dense detection responses as input, but they did not take into account in the case of sparse detection responses. Dense detection responses are difficult to obtain in complex environments all the time. In order to achieve this goal, they propose a particle-filter-based triple threshold method to build reliable trajectories. Here, they apply topic model to represent human appearance. The appearance of each person can be considered as topic distribution. Then a cost function algorithm is used to associate these trajectories in a time sliding window for final tracking results. The cost function is composed of four parts: appearance cost, motion direction cost, object size cost and distance cost. These four parts are integrated into a unified formula to build this cost function. Finally, they use three challenging datasets to evaluate the performance of the author's approach in the case of dense and sparse detection responses, respectively, and compare with state-of-the-art approaches. The results show that their approach can obtain better tracking performance than that of previous methods in both cases.