Internal Report

Segmentation and detection of individual actions from a stream of human motion is an open problem in computer vision. This paper analyzes the movements of the crucial points of the human body (hands, feet, head and center of gravity) in order to detect simple actions such as walking, jumping and displacing an object. The crucial points are considered as cooperative agents forming a team (the whole body) and their movements are analyzed at individual level and at team level. The feature extraction system is a 2D silhouette based human posture estimation technique that relies on the use of a single non calibrated camera. We also present a novel framework for online probabilistic plan recognition in cooperative multiagent systems called the MultiAgent Abstract Hidden Markov mEmory Model (M-AHMEM). The new model is the fusion of the existing single agent AHMEM and the Hierarchical Multiagent Markov Processes (HMMP), to allow multiagent policies to have internal memory that can be updated in a Markov fashion. The Rao-Blackwellized Particle Filter (RBPF) is used to perform approximate inference in the Dynamic Bayesian Network resulting in an algorithm with linear complexity O(ΣKi) with respect to Ki, the number of levels for agent i, making the system adaptable to real-time situations.