Object Tracking Algorithm Research Based On Corner Detection

Object tracking has been widely applied in many fields, such as military weapons, videosurveillance, human-computer interaction and video compression coding. Therefore,objecttracking has a very important research value and significance. To solve various problems thatexist in the object tracking, domestic and foreign researchers have proposed many trackingalgorithms. Among them, the Mean Shift algorithm has been the research focus in the objecttracking for its rigorous theory, easy implement and good tracking performance. This paperfocuses on the algorithm, analyzes the problems in object tracking applications and proposesthe solution.The Mean Shift algorithm is an efficient search for a matching tracking, and its principleis the tracking method based on statistical modeling of the characteristics of the probabilitydensity. In the tracking process, the Mean Shift algorithm first in the starting frame of thevideo sequences, target area of interest is selected through hand, and the corresponding targetmodel is established; and then in the subsequent frame, according to the Bhattacharyyacoefficient,the most similar candidate target area with the target model is found by iterativesearch.For the change appearance, noise, and partial occlusion, the Mean Shift algorithmshows good tracking performance, but when the target and background is similar in videosequences, the classic Mean Shift algorithm is difficult to distinguish between target andbackground, making the object can not accurately be tracked, even missing the object.Mean Shift tracking algorithm can not track the target accurately or even miss the target,and in this paper, the algorithm has two improvements: on the one hand, color characteristicsof corner pixels within the target area are to build the target model and candidate targetsmodel.Because the corner is the structural characteristics of the object, which makes the targetrepresentation's ability strong to effectively distinguish between target and background,thereby weakening the interference of background information. On the other hand, theiterative weights in the Mean Shift algorithm is improved and the improved iterative weightscan distinguish the major pixels and secondary pixels within the target model to assigndifferent weights, which further effectively distinguishes between target and background, andalso reduces the amount of computation, thereby improving algorithm real-time. Theexperimental results show that when the target and background is similar in video sequence,the improved algorithm shows a good robustness. Moreover, the accuracy of the algorithmand real-time have also been improved.