Template Matching Algorithm Based On Gaussian Scale-space

Vision tracking has become a hot issue of applied mathematics, computer vision and other interdisciplinary subjects, and its work is finding out the interested target in each frame of a video. Research of the field not only has profound theoretical significance, but also has wide application value such as security monitoring, human-computer interaction, intelligent transportation, and medical diagnosis, etc.Owning to relatively simple calculation and strong robustness, template matching has become commonly used vision tracking algorithm. Recently, lots of template matching algorithms were proposed. In the broad sense, they could be classified into the intensity correlation-based method and geometric transformation-based method. Among them, intensity correlation-based method was simple and highly accurate, and got better application. But the algorithm still had many disadvantages, for example, fixed template-scale, template drift, large calculation amount, etc. In order to solve these problems, the paper analyzed the existing algorithms and studies in detail. On this basic, template matching algorithm based on Gaussian scale-space was putted forward. The main content of this paper is as following:Firstly, with the Gaussian scale-space theory, the template could adaptively update according to the scale change of the target so that the algorithm could track almost of the target under no severe shape change. So the accuracy of tracking was improved.Secondly, the reasons for the template drift phenomenon were analyzed, and the solutions were put forward. The algorithm could still keep good tracking performance under occlusions and a certain degree of shape deformation.Thirdly, an improved diamond search (IDS) algorithm was proposed, which combined their own advantages of three-step search (TSS), new three-step search (NTSS) and diamond search (DS) algorithms. At the same time, IDS algorithm increased the matching speed on the basic of searching for the global optimum, and solved the large calculation amount of template matching. Experiments under complex scenes showed that the proposed template matching algorithm was a fast and robust vision tracking algorithm, could handle scale and trajectory changes of the target, occlusions, certain degrees of deformation, etc. The idea and algorithm of the paper will be applied to the Neusoft Corporation’s production research and development.