A Fast Hybrid-kernel Correlation Filter Tracking Algorithm

Abstract

With the development of computer vision and imaging technology, high frame rate cameras improve tracking performance by higher quality videos. However, recent tracking algorithms are mostly proposed for ordinary cameras, of which the processing speed is far from enough to keep along with the acquisition speed in high frame rate cameras. In this paper, a fast tracking algorithm for high frame rate cameras was proposed. The algorithm improves the kernel correlation filter method in the kernel function and feature description. For the kernel function, a hybrid-kernel function was employed by fusing local and global kernel. For feature description and dimensionality reduction, the deformable part model without image pyramid is adopted for image description due to the improvement of video quality. The proposed fast hybrid-kernel correlation filter tracking algorithm performs better for real-time target tracking in the video of high frame rate camera than other methods. The experimental results on 50-segment videos of the OTB2015 and Need for speed (240fps) database verify the effectiveness. In OTB2015, the accuracy is 74.3% with 320fps in CPU, and the accuracy and frame rate are 70.6% and 273fps, respectively in Need for Speed.