Shape of Motion and thePerception of Human GaitsJe

Researchers in computer vision have recently demonstrated several systems that interpret motion and optical ow without using a model of kinematic structure. These non-structural methods usually integrate a eld of motion into a more compact representation. In this paper, we present the design of an experiment to investigate the relationship between the non-structural shape-of-motion algorithm and human perception of gait. We take the features used by the algorithm, and use them to synthesize gait-like optical ow. A group of subjects then views the ow stimuli and records their perceptions. The motion stimuli are designed to diier in structure but have similar shape-of-motion. We wish to show that we can vary the structure of a stimulus without altering its perception, so long as we maintain the shape-of-motion. Pilot study results illustrate the experiment design. By performing this experiment to relate gait perception with gait synthesis, we are able to probe the computer vision algorithm. Recent research in computer vision shows an interest in methods for perception of human locomotion and other activities. The methods t into two broad categories: structural and non-structural. Structural methods use a model of human kinematic structure and possibly dynamics. In contrast, non-structural methods (sometimes referred to as appearance-based or model-free) avoid using such models. For example, Little and Boyd demonstrate non-structural gait recognition using shape-of-motion features 13, 6, 14] (described in Section 2). Polana and Nelson 15, 16, 17] look at global spatial distributions of motion for a gure engaged in some activity. They are able to recognize diierent activities by comparing motion statistics computed over a coarse mesh. Baumberg and Hogg 3] give a method to describe the shape of a walking human body as a function of time. In later work 4], they describe the variation in the shape over time as the changing shape of a vibrating plate. Bobick and Davis 10] describe another non-structural approach that analyzes the shape of a motion-energy image (MEI), a summation of optical ow over a sequences of images. Features that describe shapes in the MEI are used to recognize activities. A common theme in these non-structural methods is the integration of a eld of motion into a more compact representation. There is evidence in the psychophysics literature to