A graphical model for human activity recognition

Abstract

We propose a general model for visual recognition of human activities, based on a probabilistic graphical framework. The motion of each limb and the coordination between them is considered in a layered network that can represent and recognize a wide range of human activities. By using this model and a sliding window, we can recognize simultaneous activities in a continuous way. We explore two inference methods for obtaining the most probable set of activities per window: probability propagation and abduction. In contrast with the standard approach that uses several models, we use a single classifier for multiple activity recognition. We evaluated the model with real image sequences of 6 different activities performed continuously by different people. The experiments show high recall and recognition rates. © Springer-Verlag 2004.