UC San Diego

In this thesis, I address the problem of tracking multiple identical mice through severe occlusions from video of a side of their cage. A solution to this problem would greatly benefit medical research because of the key role animal testing plays in medical research. As the majority of visual tracking algorithms are intended for tracking people or cars, they are not directly applicable to the mouse tracking problem. Mice are extremely deformable, unconstrained three-dimensional objects. They have few trackable features, and their motion is extremely erratic. In addition, the mice are visually indistinguishable. Thus, to keep track of identities one must track the mice through sometimes complete occlusions. Furthermore, because of the constraints of the application, we had little control over the content of the video sequence. Thus, the bedding of the cage was hard to distinguish from the mice, and the mice were also occluded by scratches on the front of the cage. In this thesis, I break up the tracking problem into parts: defining a state representation for the mice, defining an appearance model, defining a motion model, and inferring statistics of the positions of the mice for a novel video sequence given the defined model. Because of the difficulty and uniqueness of the mouse tracking problem, it was necessary to understand, modify, combine, and invent state-of-the-art approaches to each of these parts of the tracking problem. In this thesis, I describe and motivate the solutions chosen for each of these subproblems