Independent Viewpoint Silhouette-based Human Action Modelling and Recognition

Overview & Synopsis

This paper addresses the problem of silhouette-based human action modelling and recognition independently of the camera point of view. Action recognition is carried out by comparing a 2D motion template, built from observations, with learned models of the same type captured from a wide range of viewpoints. All these 2D motion templates, are projected into a new subspace by means of the Kohonen Self Organizing feature Map (SOM). A specific SOM is trained for every action, grouping viewpoint (spatial) and movement (temporal) in a principal manifold. This approach enables the interpolation of data ”between different viewpoints” and, at the same time, to establish motion correspondences between viewpoints without considering a mapping to a complex 3D model. Every new 2D motion template gives a distance to the map, related to the probability that motion feature belongs to that particular action. Action recognition is accomplished by a Maximum Likelihood (ML) classifier over all specific-action SOMs. We demonstrate this approach on two challenging video sets: one based on real actors making 11 complex actions and another one based on virtual actors performing 20 different actions.