Oscillation Detection and Parameter-Adaptive Hedge Algorithm for Real-Time Visual Tracking

Although correlation filter-based method performs high efficiency for visual tracking, its tracking precision may be greatly degraded when occlusion occurs. To remedy this, this paper proposes a new spectrum oscillation detection algorithm and an online learning strategy for real-time tracking. Firstly, to facilitate the tracking online learning to adjust weights itself, a weighted parameter-adaptive Hedge algorithm is presented to reduce the parameters of the adjustment. Secondly, since the spectrum of the correlation filter will fluctuate when occlusion occurs, a spectrum oscillation detection algorithm is proposed to detect the frequency spectrum response at target oscillation level. Thirdly, a backtracking algorithm is proposed to predict object position when the spectrum oscillation has been detected. Finally, an update index is introduced to determine whether the current frame is updated to improve tracking accuracy and robustness. Experiments conducted on VOT2016 and OTB-2015 demonstrate the good performance of the proposed tracking method and competitive performance against the state-of-the-art tracking methods.