Incremental Learning For Visual Tracking

In the field of computer vision and pattern recognition, video tracking is a very important research direction.Although there are immense amounts of concrete research of video tracking, a lot of issues are still not solved thoroughly, such as unavailability to real-time tracking and fast moving targets tracking, inadaptability to changes on the surface of the target object or on light condition, and inadaptability to target object obscured.In the first place, this text presents the background and current situation of target recognition and target tracking, which have significant application value and abundant research history. Afterwards it presents the deficiency of existing algorithms.Then the text presents basic knowledge of template updating based on incremental principal component analysis, in which we analyzed the basic principle of principal component analysis first and then listed singular value decomposition, described mainly the method of updating sample characteristics and sample mean of which the sequence Carlo transformation is the main part, and listed the distinction in calculation between this technology with the traditional one.Besides, the text explains the technological process of tracking target by video. We stated the basic procedure of tracking target objects in the algorithm used in this paper, and then it leads to elementary particle filter and we introduced the theory of Monte Carlo method and its application in target tracking.On the basis of above research, the text puts forward a kind of integrated video tracking algorithm in the framework of improving incremental principal component analysis updating template. We improved incremental characteristics and average method. And lead in the concept of forgetting factor when we are dealing with the proportion of new observed data and previous observations. In order to make the template adapt to the change of the target surface better, we adjusted flexibly the proportion of the old and new data in updating template. What’s more, we improved the dynamic model and observation model, and reselected distance metrics.Experimental results show that this algorithm could complete the tracking commendably whenever the target object changes its post, ratio or illumination.