Context Adaptive Visual Tracker in Surveillance Networks

Abstract

CNN-based visual trackers has been successfully applied to surveillance networks. Some trackers apply sliding-window method to generate candidate samples which is the input of network. However, some candidate samples containing too much background regions are mistakenly used for target tracking, which leads to a drift problem. To mitigate this problem, we propose a novel Context Adaptive Visual tracker (CAVT), which discards the patches containing too much background regions and constructs a robust appearance model of tracking targets. The proposed method first formulates a weighted similarity function to construct a pure target region. The pure target region and the surrounding area of the bounding box are used as a target prior and a background prior, respectively. Then the method exploits both the target prior and background prior to distinguish target and background regions from the bounding box. Experiments on a challenging benchmark OTB demonstrate that the proposed CAVT algorithm performs favorably compared to several state-of-the-art methods.