Research On Mouse Tracking Based On Particle Filtering

Object tracking based on image sequence, which is a significant and challenging research topic in computer vision, has been widely and deeply studied. However, the majority of the research works are focused on the tracking of man as well as maneuvering target such as cars and planes for a long time. It is not exaggerate at all to assert that the tracking of animals in video sequences is full of challenge. Especially, the mouse targets bring enormous difficulties in tracking, which are the varieties of shapes, irregularities of motion, high frequent and serious occluding, as well as high similarities between targets leading to difficulties to distinguish them.Based on the framework of particle filter, this paper proposed and implemented different tracking algorithms for single mouse target and multi mice targets separately, which are as follows:We combined the weighted color histogram, which is considered as observation model, into the framework of particle filter and realized the tracking of single mouse target. The proposed method can track mouse target successfully when there are a bit large number of particles.In order to track multi mouse targets at the same time, this paper first analyzed the dynamic model of mouse motion, and then discussed the dynamic change regularities of position as well as shape. Next, a 2nd Auto Regressive Process (ARP) is proposed to describe the position transitions of mouse targets and finally constrained Brown motion is utilized to describe the dynamic change of mouse shapes.To build the appearance model of mouse target in video sequences, this paper studied a general method to retrieve image observation under the Bayesian framework. First, the image is partitioned into numerous grids, and the image observation is assumed as independent between different grids. Then the image observation of a whole image can be configured based on the configuration of image observation in each grid. Finally, based on target dynamic model as well as image observation model, a tracking algorithm for multi mouse targets is proposed together with particle filter.