Adaptive Saliency Aware Spatially-Regularized Correlation Filter for Object Tracking

Abstract

Since correlation filter (CF) can achieve a relative tradeoff between accuracy and speed, it has become a mainstream object tracking algorithm in the past decade. However, CF extends the training sample by circular shifts of a given object patch which consequently produces inevitable boundary effects. Such effects will seriously affect the tracking results, especially in challenging situations, such as occlusion, scale variation and background clutters. In order to solve them, we propose an adaptive saliency aware spatially-regularized correlation filter (ASARCF). First, an adaptive spatial regularization component is incorporated to penalize the CFs values of the non-object regions. Second, a saliency feature map of the object is integrated into CF to enhance the ability of the tracker to extract a tracking object from complex backgrounds. Third, to adapt the filter to the rapid variation of the object, we modify the learning rate of filter template by adaptive adjustment during tracking. Experimental results demonstrate that our proposed ASARCF shows competitive performance among state-of-the-art object tracking methods.