Mining motion patterns using color motion map clustering

Automatically extracting previously unknown behavior patterns from videos that track animals with various physical conditions can accelerate our understanding of animal behaviors and their influential factors, resulting in major medical and economic benefits. Unfortunately, extracting behavior patterns from videos recordings remains as a very challenging task due to their extensive duration and the unstructured natures. This task is further complicated in a completely darken animal cage with inconsistent infrared lighting, moving reflections, or other cage debris such as the cage bedding. In this research, we propose a new motion model that enables us to measure the similarities among different animal movements in high precision so a clustering method can correctly separate recurring movements from infrequent random movements. More specifically, our model first transforms the spatial and temporal features of animal movements into a sequence of color images, referred to as color motion maps (CMMs). The task of mining recurring behavior patterns is then reduced to clustering similar color images in a database. We will use a real infrared video to demonstrate the capability of our model in capturing distinguished but brief animal movements that are embedded within a sequence of other animal movements.