CNN-detector-based multiple homogeneous objects tracking under stochastic wide-range occlusions
Yan Li, Xianbin Cao, Junying Liu, Baochang Zhang · 2017
In this paper, we address a specific, challenging multi-object tracking (MOT) problem, in which targets are in small sizes and demonstrate highly homogeneous appearance in the vast scene with stochastic huge occlusions. To tackle those challenges, we propose a novel method that integrates realtime detection and tacking in one single learning framework. Specifically, we formulate MOT as a clustering problem in a parameterized semi-metric space, where we introduce a CNN detector and a space mapping to partition (SMP) tracker. In our method, the length of the association temporal window is automatically and adaptively selected according to the quality of detection responses, which enables robust detection and tracking. Extensive experiments have been conducted on a newly collected air-traffic-control (ATC) dataset to validate the effectiveness of the proposed method. Results have shown that our method can achieve high tracking performance and substantially exceeds state-of-the-art methods by large margins.