An Approach to Overcome Occlusions in Visual Tracking: By Occlusion Estimating Agency and Self-Adapting Learning Rate for Filter's Training

Abstract

Visual tracking methods have been successful in recent years. Correlation filter (CF) based methods significantly advanced state-of-the-art tracking. The advancement in CF tracking performance is predominantly attributed to powerful features and sophisticated online learning formulations. However, there would be trouble if the tracker indiscriminately learned samples. Particularly, when the target is severely occluded or out-of-view, the tracker will continuously learn the wrong information, resulting target loss in the following frames. In this study, aiming to avoid incorrect training when occlusions occur, we propose a regional color histogram-based occlusion estimating agency (RCHBOEA), which estimates the occlusion level and then instructs, based on the result, the tracker to work in one of two modes: normal or lost. In the normal mode, an occlusion level-based self-adopting learning rate is used for tracker training. In the lost mode, the tracker pauses its training and conducts a search and recapture strategy on a wider searching area. Our method can easily complement CF-based trackers. In our experiments, we employed four CF-based trackers as a baseline: discriminative CFs (DCF), kernelized CFs (KCF), background-aware CFs (BACF), and efficient convolution operators for tracking: hand-crafted feature version (ECO_HC). We performed extensive experiments on the standard benchmarks: VIVID, OTB50, and OTB100. The results demonstrated that combined with RCHBOEA, the trackers achieved a remarkable improvement.