The Study Of Human Action Recognition Method Of Video Data

Human action recognition is the foundation of human behavior comprehension, it is of great significance and can be used in lots of real-world applications. Human action recognition commonly include following steps:motion detection, feature extraction, classification. In this paper, we first introduce some widely used motion detection algorithms, including Gaussian Mixture Model, ViBe, Optical Flow.We use two methods on human action recognition. The first one is3D Zernike moment, this is a method based on spatio-temporal feature extraction, it treat the human action as space time shape, then calculate the3D Zernike moment of it as the global feature of the action, classify the features use SVM classifier at last.The second method is to use sparse representation in human action recognition. Unlike classic spatio-temporal feature extraction methods, it transform the original data to another representation space. We design two experiments to verify it is effective or not. In the first experiment we use sparse representation as a feature extraction method, first we transform the original video data into sparse coefficient, then classify these coefficients instead. The second experiment is based on the SRC(sparse representation based classification) method. We treat all frames in the training data as the over complete dictionary, use this dictionary to sparse represent the test video frames, then sum up the sparse coefficient like in motion history image method to obtain the final feature of the test video. The result of the experiments show that when use sparse representation as a feature extraction method, the correct rate is lower than state-of-the-art, thus this method still need improvement, the result of the second experiment outperforms some space time feature extraction based methods.