A Robust Human Action Recognition System Using Single Camera

This paper describes a system for action recognition with a single camera. Firstly, we use a two-layered background subtraction, which is based on both chromaticity and gradient, to extract human contours from the frame sequence captured by camera. This subtraction helps us remove shadows from foreground and get a good contour for recognition. Then we parameterize a human posture with a model called star skeleton. Each posture is represented as five vectors by this model. To classify every posture into a symbol, we trained a classifier by SVM with a set of exemplars. This classifier calculates a symbol according to the star skeleton. So it transfers the frame sequence into a symbol sequence. Finally, an action is recognized from the symbol sequence with a string matching. Our system can be regarded as a new way for interaction between human and computer. It achieves a high and stable recognition rate in complex environment, and the experimental results show the promising effectiveness.