High-Speed Tracking With Kernelized Correlation Filters Algorithm Research And System Implementation

Visual tracking is a hot issue within the field of computer vision. Its major task is to get trajectory of the interested target from video sequences. Visual tracking technology has important significance in video retrieval, human-computer interaction, automated monitoring and defense combat.The performance of target tracking algorithm is generally measured by tracking accuracy(the precision rate) and tracking speed(the frame rate). In military application, for fast moving military targets tracking, it needs especially to meet the real-time requirements. This paper mainly studies the tracking algorithm KCF(Kernelized Correlation Filters) which is based on kernel correlation filtering, and has the advantage of high speed.KCF is based on the idea of cyclic shift and apply Discrete Fourier Transform to ridge regression, simplifying the solution of the regression problems, improving the tracking speed and getting the jump on the current algorithm for optimal performance in terms of accuracy and frame rate, such as TLD and Struck. However, there are still some defects in the algorithm, such as unable to adapt to changes in the target scale, lack of strategies to resolve the problems of target missing.In this paper, we first study the KCF and optimize for the shortcomings of the algorithm. Then, using C ++ and Open CV, we build a high-speed target tracking system which KCF is applied to, achieving good tracking performance.The main work is in the following areas:(1) To resolve the problems of target missing in KCF algorithm, we proposed an optimization algorithm: determine whether the target is lost, according to the response value between current frame and the first frame, and the response value between current frame and the model. Redetect targets using frame differential method, K-Means, response values between frames, reinitialize the track model and continue to tracking process. Experiments show that the optimization algorithm has good robustness for videos with shade problems.(2) To deal that KCF algorithm has not optimized the parameters of HOG features, we analyze the effects of different HOG feature parameters on the tracking accuracy and tracking speed, and optimize the parameters selected by experiment. The accuracy of the optimized algorithm based on linear kernel increased 1%, the frame rate increased by 50%. The accuracy of the optimized algorithm based on gaussian kernel increased 2%, the frame rate increased by 65%.(3) Described the implementation of high-speed target tracking system and analyze tracking performance of the system by experiment. Using the 50 videos In CVPR 2013 as a benchmark, the 20-pixel precision as an evaluation criterion, the average accuracy of the tracking system reached 74.7%, and the average frame rate reached 466 per second. The system basically meets the requirements of high-speed real-time tracking.