Behavior recognition in laboratory mice using HFR video analysis

Abstract-In this paper, we propose an automatic behavior recognition algorithm for laboratory mice to determine their typical model behaviors and hence detect the degrees of diseases via animal testings for behavioral pharmacology. The algorithm can quantify several model behaviors of mice by detecting the repetitive motions of fore or hind limbs at several hertz; these motions, too rapid for the naked eye and the NTSC cameras, are captured on a high-frame-rate video. Even when a mouse changes its position and orientation, the algorithm can always detect these repetitive motions as periodic frame differential image features on four segmented regions-head, left side, right side, and tail-because the silhouette image of a mouse is converted in the polar coordinate system as its shift and rotation-invariant shape by detecting the positions of its head and tail. The effectiveness of the algorithm is evaluated for six typical model behaviors, including scratching and grooming, by analyzing the experimental results for several laboratory mice, obtained by capturing long-term videos at 240 fps. We have provided the detection/correction ratios and compared the time durations of the model behaviors of these mice.