Visual Tracking via Adaptive Context-Aware Correlation Filter

Abstract

Correlation Filter (CF) based trackers have recently achieved impressive performance. In this paper, we propose an adaptive way to integrate context information into CF trackers other than roughly giving the same weight. Moreover, in order to solve the challenging problem of target occlusion, a new anti-occlusion criterion is implied. First of all, we use the Kalman filter to estimate the target motion state for the purpose of predicting the motion direction. Second, context information, four direction regions around the target are extracted as background samples. During the filter training process, larger weight is assigned to the background samples which corresponding to the motion direction. Then a new occlusion criterion named APCE is implied while the model is updated. Furthermore, extensive experiments on OTB-2013 dataset show that our approach outperforms most mainstream tracking algorithms in terms of precision and success rate.