Adaptive Label-Constrained Correlation Filter for UAV Tracking

Abstract

Traditional discriminative correlation filter (DCF) tracking algorithms always use ideal Gaussian functions as labels to train filters, which has reached promising performance in usual scenarios. However, due to challenges such as camera motion, occlusion and similar targets appearing frequently in unmanned aerial vehicle (UAV) tracking scenarios, these trackers using the identical and fixed labels often lead to over-fitting and model degradation and therefore perform poorly in UAV tracking. Accordingly, we present a new framework named Adaptive Label-Constrained Correlation Filter (LCCF) to adaptively construct a more realistic label function for each frame. Specifically, we propose adaptive label constrain regularization terms to assist in the construction of the desired realistic label function. In addition, we introduce an additional temporal regularization term to ensure the temporal consistency, thus avoiding using an additional fixed learning rate. Broad tests on multiple challenging UAV datasets have strongly established the comparative advantage of LCCF over deep and DCF methods. Moreover, LCCF fulfills the real-time tracking requirements with a tracking speed of 43 FPS. Remarkably, our approach delivers new best performance on VisDrone.