Facing Completely Occluded Short-Term Tracking Based on Correlation Filters
Yuanming Zhang, Huihui Pan, Jue Wang, Weichao Sun
Abstract
Accurate occlusion tracking is a difficult research in vision-based measurement. Most state-of-the-art approaches exploit suppression of templates learning of filters to avoid occlusion pollution. Such suppression strategies retard in the recognition of target obstruction and lack state processing during the occlusion phase. This paper explores the precise and robust handling of the complete occlusion problem in the framework of discriminative correlation filters. We propose a generative association approach by utilizing raw images of pixels as well as adjacent trajectories for generating robust target descriptors and observable cumulative sequences. The occlusion of the target is identified through the descriptors matching over the time-series, and estimates the transfer of the state by the generating function. Compared to suppression methods, our approach develops the apparent characteristics and cumulative state of the target, not relying on filter templates. This achieves a maximum gain of 5.0 percent of area under the curve precision on the OTB benchmark. Comprehensive evaluations are performed on six datasets: UAV123, TB50, OTB100, LaSOT, VOT2018ST, and VOT2019ST. Our method establishes a new record score in distance precision on the TB50, by exceeding the other 19 state-of-the-art trackers. Also, we reach the state-of-the-art ranks on both the VOT2018ST and VOT2019ST datasets.