Research On Target Tracking Method Based On Kernelized Correlation Filters

With the rapid development of computer performance, image processing and machine learning technology, computers are replacing human to process data gradually. How to manipulate the large computational complexity of the video and analysis image information by computers is becoming the key research field of artificial intelligence. As the visual target tracking is widely used in human-computer interaction, video surveillance, national defense security and other fields, the system is facing more and more challenges, such as illumination variation, deformation, scaling, motion blur, occlusion and out of view. Therefore, it’s a challenging task to research on robust practical target tracking algorithm.In this article, we make deep and systematic researches on kernelized correlation filters. To address the problems in target tracking, some effective improvement methods are proposed on the basis of previous research. The specific research work summarized as follows:Firstly, in order to adapt to out of view and occlusion of target during tracking, a kernelized correlation filters based on sparse representation is proposed. In the tracking framework of kernelized correlation filters, sparse representation is used as the template of target tracking. The template updates while the tracking result is reliable. The proposed update mechanism can effectively avoid model error updating of kernelized correlation filters.Secondly, to improve the tracking accuracy of target tracking algorithm further, a Coarse-to-Fine framework of kernelized correlation filters is proposed. Both translation template and fixed template are constructed for target during tracking. More candidate areas are positioned by translation template, the fixed template is used to provide exact orientation with candidate areas. The experiments prove that the proposed algorithm improves tracking accuracy significantly and it can effectively cope with the challenge of occlusion, deformation and rotation.Thirdly, a fast feature extraction method adapted to scale change is proposed. In the tracking framework of kernelized correlation filters, this paper proposes a rapid method of feature extraction method, which is extracted the space combination feature of gradient histogram. This method can effectively avoid repetitious feature extraction procession, and reduce the redundant data but not key information. Experiment proves that this algorithm can quickly extract the feature of image, and adapt to the scale change of target.Concentrating on problems of target tracking, kernelized correlation filters is thorough researched. Corresponding improval methods for different questions are being put forward. Sequences of experiment are selected from target tracking field authority standard video library(Visual Tracker Benchmark) as a data source. Several state-of-the-art target tracking algorithms are used to compare with the algorithms we proposed. Experiments show that the proposed algorithms have good performance to deal with a variety of difficulties, including out of view, occlusion, deformation and scale variations.