Three Brown Mice: See How They Run DRAFT: Do Not Distribute Kristin Branson and Serge Belongie Vincent Rabaud Dept. of Computer Science and Engineering Coauthor Department U.C. San Diego Coauthor Institute La Jolla, CA 92093 City, STATE zipcode Abstract We address the problem of tracking multiple, identical, nonrigid moving targets through occlusion for purposes of rodent surveillance from a side view. Automated be- havior analysis of individual mice promises to improve animal care and data collection in medical research. In our experiments, we consider the case of three brown mice that repeatedly occlude one another and have no stable trackable features. Our proposed algorithm com- putes and incorporates a hint of the future location of the target into layer-based affine optical flow estima- tion. The hint is based on the estimated correspon- dences between mice in different frames derived from a depth ordering heuristic. Our approach is simple, efficient, and does not require a manually constructed mouse template. We demonstrate encouraging results on a challenging test sequence containing multiple in- stances of severe occlusion. 1. Introduction In this paper we consider the problem of multiple ob- ject tracking in the case where the objects are in- distinguishable and prone to occluding one another. Recently a number of works have appeared that ad- dress the problem of multiple object (or blob) tracking [5, 1, 10]. Many of these approaches leverage object- specific appearance models such as color histograms [1]. Most closely related to our problem setting is that for which the BraMBLe algorithm was designed [5]; this reference also provides a thorough review of relevant visual tracking methods. In this work, the authors use a particle filter to track multiple blobs (viz. people) from a ceiling-mounted hallway camera. The principal failure mode of their system is that of blob labels (i.e. identities) getting switched when one object passes in front of another. The authors suggest that the use of individual foreground models for each object could be used to solve this problem. In our setting, however, the Figure 1: Still frame captured from video sequence of three mice (240 × 360 pixels). The metal prism at the top of the cage contains food pellets and a water bottle. It also prevents the use of an overhead mounted cam- era. The bedding on the floor of the cage is the only dynamic part of the background, other than reflections. objects we wish to track are mice, and they all have the same appearance; see Figure 1. Aside from the challenge of tracking identical targets through occlusion, our problem setting presents a num- ber of other difficulties. The objects we wish to track have little or no trackable features (in the sense of [9]) that last for more than a few frames. Additionally, the motion of the objects is relatively erratic compared to a car driving past an occluding signpost, for example. On the other hand, we benefit from a number of sim- plifying assumptions, e.g. that the number of objects does not change, the illumination is relatively constant, and the camera is stationary. Because of these simplifying assumptions, the sub- problems of foreground/background classification and tracking for separated mice (mice that are neither oc- cluded or occluding) are adequately addressed by many existing algorithms. The subproblem remaining is to 1