关键词: 目标跟踪, 上下文信息, 显著特征, 特征增强, 深度学习

Abstract: Abstract: Making full use of the temporal and spatial context information in the video can significantly improve the performance of object tracking. However, most of the current algorithms based on deep learning only use the feature and information of the current frame to locate the object, the tracking object is apt to interference from the similar object nearby, as well as the potential error will be cumulatively introduced into tracking process. Therefore, it is necessary to utilize the spatio-temporal context information of the object of interest in the previous and latter frame of the video. In order to retain temporal and spatial context information, a short-term memory storage pool is introduced to store features of the previous frames. At the same time, we propose an appearance saliency feature boosting module (ASBM), which not only enhances the saliency feature of the tracking object, but also restrains the interference from similar object around the tracking object. In this paper, we propose an object tracking algorithm based on spatio-temporal context information enhancement. We have designed experiments on four data sets, which are VOT2016, VOT2018, DAVIS-2016 and DAVIS-2017, to verify and compare the performance of our method. Compared with the SiamMask algorithm, the accuracy and average overlap rate of the proposed method in this paper are increased by 4% and 2%, respectively. The accuracy, robustness, and average overlap rate are improved by 3.7%, 4.2% and 1%. In the regional similarity of DAVIS-2016 and DAVIS-2017, the contour accuracy has been improved by 2% and 1%, respectively.

Key words: object tracking, contextual information, salient features, feature enhancement, deep learning