Learning Rate Adaptive Kernel Correlation Filter Tracking Algorithm Combined with Multi-feature Fusion

Abstract

The tracking approaches based on correlation filter have shown their particular importance due to the performance superiority from the perspectives of accuracy and speed. However, some issues of these tracking approaches are still to be addressed, such as target scale variation and occlusion environment. In this paper, we propose a learning rate adaptive kernel correlation filter tracking algorithm combined with multi-feature fusion to address the issues presented above, whereas existing KCF only uses single feature. In the proposed approach, color attributes and histogram of oriented gradient are combined for the subsequent dimensionality reduction by principal components analysis for the precise feature presentation of targets. Besides, we employ average peak-to correlation energy (APCE) in the update of our model learning rate for solving the problem of serious occlusion of the target. Our proposed tracker keeps a good performance for the scale variation by using an accurate scale estimation method. The experimental results on OTB50[1] dataset show that a significant performance improvement is obtained by our approach, compared with other methods based on correlation filter trackers, where the tracking precision is 0.84, success rate is 0.82. At the same time, it has a real-time tracking speed of 40 frame/s.