The Study Of Vision-based Human Action Recognition

The Study of Vision-based Human Action Recognition is one of hot researches in the computer vision field. It has a broad prospect of application and economic value in intelligent surveillance, human movement analysis, advanced user interface, virtual reality and video coding and transmission areas, so attracts the researchers'attention widely. Due to the complexity of human movement itself and the diversity of outside environment, make human action recognition having certain challenging. At present the research of human action recognition is still in the primary stage and remains to be further research and study.This paper launches the research around the related problems of vision-based human action recognition, mainly including moving object extraction, feature extraction and algorithm of human action recognition three aspects. Moving object extraction is the research foundation of vision-based human action recognition. First of all, making corresponding pretreatment for the videos such as separating frame and making gray, extracting human binary image by background difference method and using mathematical morphology processing the holes and noise point existed in binary images. With the high of minimum circumscribed rectangular as side length, extracting the Region of Interest (ROI). On this basis, this paper takes feature extraction and algorithm of human action recognition as the key researches. This paper extracts human action feature by combining shape characteristics and movement characteristics. Using the Principal Component Analysis (PCA) to reduce the dimension of ROI and getting the shape characteristics. At the same time, calculating the height-width ratio, posture change velocity and movement rate as the movement characteristics. Then, adopting RBF neural network algorithm based on clustering to classify and recognize human action, and using cross-validation method to evaluate the algorithm.This article selects five kinds of human actions in the Weizmann database as the data set on the human action recognition system for experimental simulation in the environment of MATLAB R2008b, and proves the validity of the algorithm by the analysis of experimental results.