Modeling trajectories of free moving objects with smooth flow fields

Abstract

We present a new method for smooth global analysis and representation of image-based motion data with topological methods for flow fields, useful for further activity analysis. A video sensor captures motion of subjects and yields discrete samples of velocity vectors in the image domain. We show how to construct a smooth, continuous, globally defined bidirectional flow field which approximates the directional part of given samples. We then apply a topological analysis of this continuous flow field which yields a segmentation into regions of similar flow behavior, i.e., regions of similar motion. This segmentation (topology) can be seen as global model of typical motions and used for further analysis like detecting atypical motions. We tested our method empirically and provide results for two different scenarios with human motion and traffic motion, respectively.