Research Of Human Action Recognition Algorithm In The Video

Action recognition based on video is an hot research in the field of computer vision, which aims to analysis and recognize the human action in video automatically by the computer. It can be applied to intelligent video surveillance and human-computer interaction. In the field of action recognition, motion description plays an important role, and the covariance matrix meets demand appropriately, which reflect the statistical change of image features and measure the change magnitude of actions. In this paper, we summarize the research about action recognition, and research a new algorithm about action recognition based on the covariance matrix. This paper includes three parts such as global feature, local feature, representation and classification of actions, and then details are follows:(1)We research an algorithm of extracting global feature about actions. The action video can be divided into several overlapping segments, and we will calculate the covariance matrix of each segment with human’s HOF features, from which the global feature can be extracted. After the global features’ sparse code, we will get their sparse coefficient histogram as a feature vector of video.(2)We research an algorithm of extracting local spatial-temporal feature. We can get a lot of cuboids by dense sampling. The next step is to calculate covariance matrix of cuboids and extract local spatial-temporal from the covariance matrix. To get it, we have two methods, one is based on a variety of image features, and another is based on single image feature. We will compare those two methods under the same experiment. The following experiment indicates that the former has higher recognition rate. In addition we will also research the different descriptors based on combinations of different image features.(3)We research the method of representation and classification about actions. First we learn an over-complete dictionary with the descriptors of actions in video. Second we encode all of the descriptors to get their code coefficients. We can obtain feature vector of the motion video after pooling with spatial pyramid model. Last we use SVM to recognize feature vector of human action, and compare the recognition rates of different kernels.The main tool of our algorithm is MATLAB R2012 a, and databases for test are KTH and UCF Sports. Our algorithm of local feature based on covariance matrix has a high level from the test results of the algorithm.