Research On Basic Human Action Recognition By Multiple Feature Fusion

With the explosive growth of videos on Internet, recognition on human action by computer vision methods has becoming a hot topic. Wide range of application areas are the most attractive aspects of this technology, such as military, medicine, security and many other aspects. The growing demand and applications have promoted a rapid development of this technology in a short period of time. In recent years, the combination with disciplines such as machine learning and data mining has made the research deeper and more comprehensive. The process generally consists of target detection tracking、feature representation and action recognition. But most of the current recognition can only be based simple action, and many methods can obtain better recognition results only on some particular research areas.This study can mainly be applied to basic human action. The purpose of this study is to detect the moving body in the video firstly. Then multiple features have been calculated to represent the moving target. And ultimately the human action in the video has been classified and identified. This article has given a detailed description on recent research issues and difficult problems of human action recognition. Some general target detection methods like background subtraction have been introduced and compared. Moreover, the combination of the optical flow algorithm with ViBe algorithm has been adopted to estimate the moving target. And the fusion of multiple features has been applied to describe the movement information in the video. The detailed process of this experiment shows below. In the target detection phase, the optical flow algorithm has been used to estimate the moving area of the target and eliminate the influential areas from noises. Then the ViBe algorithm has been selected to segment the target accurately from the background on the basis of its property of modeling background by sample and random aggregation. In the feature extraction phase, multiple features like SIFT descriptors and Fourier descriptors have been calculated after the preprocessing. And the feature fusion work has finished by cumulative histogram. At last features are combined with artificial neural network to perform classification.The research selects the KTH video data sets which contain600basic human actions as the research objective. The training of feature points and classification has been performed with Microsoft Visual Studio2010and OpenCV computer vision library. Good recognition results prove the effectiveness of methods in this research.