A View-Based Real-Time Human Action Recognition System as an Interface for Human Computer Interaction

Abstract

This paper describes a real-time human action recognition system that can track multiple persons and recognize distinct human actions through image sequences acquired from a single fixed camera. In particular, when given an image, the system segments blobs by using the Mixture of Gaussians algorithm with a hierarchical data structure. In addition, the system tracks people by estimating the state to which each blob belongs and assigning people according to its state. We then make motion history images for tracked people and recognize actions by using a multi-layer perceptron. The results confirm that we achieved a high recognition rate for the five actions of walking, running, sitting, standing, and falling though each subject performed each action in a slightly different manner. The results also confirm that the proposed system can cope in real time with multiple persons.