Learning Aberrance Repressed and Temporal Regularized Correlation Filters for Visual Tracking

Abstract

Recently, in visual tracking field, discriminative correlation filter (DCF) methods have achieved competitive performances. However, the target template of the DCF may be polluted by the deformation, illumination variation, and scale variation, resulting in drift and tracking failure. To solve this problem, we propose an aberrance repressed and temporal regularized correlation filter. A novel tracking algorithm is presented by introducing a temporal regularization into ARCF tracker, which can not only let the filter template retain the historical information for filter learning, but also achieve a long-time and high precision model, compared with ARCF, which happened in large complex variations. Furthermore, hand-crafted and deep features are combined to achieve superior feature representation. Finally, the extensive experiments are conducted on OTB2015, VOT2018, and VOT2016. Specifically, compared with popular trackers, our final tracker algorithm performs well and obtains AUC score of 69.6% and DP score of 92.2% on OTB2015, besides, our tracker also works well compared with other related methods and achieves EAO score of 0.422 on VOT-2016.