Behavioral quantification of freely moving mice

Movies of freely moving animals contain rich information about an animal’s behavior. However, unbiased quantification of the behavior remains challenging. In our work, we investigate two different strategies for behavioral classification used in previous studies: First, we perform wavelet transformations of the distances between body parts, followed by unsupervised clustering of the data in frequency space to identify distinct behaviors. As a second approach, we identify change points in the temporal dynamics, and locally fit the data to autoregressive models, which also can be clustered to assign behaviors. By comparing these two different approaches, we can identify their strengths and limitations and create a more holistic picture of animal behavior. We use our findings to explore the behavioral repertoire of mice and identify differences between the open field behavior of mice with and without a cerebellar perturbation.