Distribution To Match The Distance Learning In The Snake Model

Tracking of moving objects in image sequences is one of the most important research topics in computer vision, which is an active field in academia. It also has very important applications in many fields. In the past several decades, active contour model (snake model) has been attracting more and more researchers as a popular tool for object tracking, and it has made a great success. However, there are still many theoretical problems with this method that need to be solved, such as initialization problem, parameter sensitivity problem, global optimum problem and so on. Thus, there is still a lot of work to do about this topic.This thesis focuses on the subject of active contours for tracking distributions, and the main work we do is as follows:(1) Quadratic form distance is adopted to define the similarity between two distributions. This measure has taken non-corresponding bins into consideration. It can be shown that with a better choice of similarity matrix the quadratic form distance is indeed a metric. Generally, researchers assign constant value to every element of similarity matrix. However this constant similarity matrix doesn't fit all the image sequences. Thus, in order to enhance robustness of our approach, we propose an adaptive similarity matrix obtained by the learning of priori knowledge and update it online. So we can minimize the distance between similar distributions and maximize the distance between dissimilar distributions.(2) We choose gray histogram, color histogram and local binary pattern as image features, and they are based on photometric variable, such as intensity, color, and texture. Compared with geometric variable (such as edges), from the global perspective, they discribe the appearance characteristics of interesting object more exactly.(3) We add shape energy and penalization energy into energy function to simplify the experiment and improve the tracking stability of the algorithm. Finally, by using Euler equation, we derive the evolution equation for the level set function.Compared with the traditional active contours for tracking distributions, the method proposed by this paper can effectively measure the distance between distributions because it uses quadratic form distance as a metric, learns similarity matrix priori and updates it online. So it is more suitable for tracking. The experimental results verify the effectiveness of the models.