Multi-kernel support correlation filters with temporal filtering constraint for object tracking

This paper proposes the adaptive multi-kernel support correlation filters with hedge parameter strategy and temporal filtering constraint for real-time tracking. In order to fuse the excellent properties of various views that characterize the object robust appearance accurately, support correlation filtering responses from multiple kernels can be adaptively integrated into one strong and accurate filtering response map by hedge parameter strategy in a parallel way. It absorbs the strongly discriminative ability from different feature-based support correlation filters, which tolerate sampling outliers of circulant structures with the help of SVM learning way. Also, it exploits the intense information of multi-view appearance representations which guarantee the fusion of reliable correlation filtering maps with reasonable parameters. Meanwhile, with the temporal filtering constraint to maintain historical appearance characteristics, alternating fixed-point algorithm improves complementary memory-updated model that keeps the stability of tracking process continuously and alleviates the target drifting situation for each support correlation filter. Experimental results demonstrate that the proposed approach achieves favorable performance on multiple dynamic scenes.