MP-LN: motion state prediction and localization network for visual object tracking

摘要

Abstract Visual object tracking is an important topic in computer vision, where the methods extracting features from the appearance of the object have made a significant progress. However, occlusion and rapid motion cause an incomplete appearance of the object and an incorrect search area in a complex scene, which limits the precision of object localization. In this paper, we propose a novel motion state prediction and localization network, named MP-LN, for visual object tracking, which predicts and translates a reasonable search area depending on the continuous motion state. Specially, we design a motion state prediction model based on the reinforcement learning, which adopts the policy gradients to estimate the target motion and incorporates rewards to enhance the back-propagation of errors for more accurate motion state. After that, we utilize an iterative localization to fine-tune the identification of the target location, reducing the response suppression. Extensive experiments and results demonstrate the effectiveness and advancement of the proposed method on six challenging tracking datasets, DTB70, UAVDT, UAV123, LaSOT, GOT-10k, and OTB2015.