Abstract
Animals can perform complex and purposeful behaviors by executing simpler movements in flexible sequences. It is particularly challenging to analyze behavior sequences when they are highly variable, as is the case in language production, certain types of birdsong and, as in our experiments, flies grooming. High sequence variability necessitates rigorous quantification of large amounts of data to identify organizational principles and temporal structure of such behavior. To cope with large amounts of data, and minimize human effort and subjective bias, researchers often use automatic behavior recognition software. Our standard grooming assay involves coating flies in dust and videotaping them as they groom to remove it. The flies move freely and so perform the same movements in various orientations. As the dust is removed, their appearance changes. These conditions make it difficult to rely on precise body alignment and anatomical landmarks such as eyes or legs and thus present challenges to existing behavior classification software. Human observers use speed, location, and shape of the movements as the diagnostic features of particular grooming actions. We applied this intuition to design a new automatic behavior recognition system (ABRS) based on spatiotemporal features in the video data, heavily weighted for temporal dynamics and invariant to the animal's position and orientation in the scene. We use these spatiotemporal features in two steps of supervised classification that reflect two time-scales at which the behavior is structured. As a proof of principle, we show results from quantification and analysis of a large data set of stimulus-induced fly grooming behaviors that would have been difficult to assess in a smaller dataset of human-annotated ethograms. While we developed and validated this approach to analyze fly grooming behavior, we propose that the strategy of combining alignment-invariant features and multi-timescale analysis may be generally useful for movement-based classification of behavior from video data.