Discriminative Sparse Coding Methods For Human Action Recognition In Video

With the development of video capturing technology, there appear more and more video data which contains more and more sufficient information. As the increasing of video data, analysising and understanding these videos have become quite urgent tasks. However, it is impossible to artificially understand the content contained in these videos. Therefore, we must develop new approaches to video analysis and understanding with the help of computer. It is at the beginning stage of the research on this topic and the topic is catching more and more attensions.Human action recognition is a conventional video understanding problem. It determines the mode of human action by analysising the relevance between images in video as well as their visual appearance. For example, one usually wants to distinguish a walking man from another running one. Since human action recognition greatly helps understanding and processing the video data, it has been widely applied to video surveillance, virtual reality, video based content retrieval and so on. Thus, it is practically valuable to conduct research on human action recognition. In addition, human action recognition has been a very challenging problem in computer vision because there are occlusions between different individuals as well as different parts of a body. Therefore, it is also theoretically important to conduct research on human action recognition. This thesis focus on the human action recognition problem and the main works include:Firstly, we compare and analysis the existing video feature extracting methods. By using MoSIFT to extract the temporal-spatial local features and introducing average pooling to obtain the final video level feature, we preserve most information contained in the descriptor of interesting points extracted from videos.Secondly, we propose a sparse discriminative analysis method based on the theory of sparse coding. It computes the dictionary by using the dictionary learning method and computes the sparse coefficients by using the sparse coding methods, and then extracts the main discriminative features by using linear discriminative analysis, and thus improves the recognition accuracy.Thirdly, we propose a discriminative sparse coding method by combining sparse coding and linear discriminative analysis. The new method introduces the discriminate power in low-dimensional space during the dictionary learning, and thus further improves the human action performance. In addition, we developed a novel online discriminative dictionary learning algorithm to preserve the discriminative information contained in the training samples.Finally, the experimental results show that: (1) The proposed sparse discriminative analysis method slightly outperforms the state-of-the-art method in terms of accuracy; and (2) By directly incorporating discriminate power during the learning procedure, the proposed discriminative sparse coding method preserves as more discriminative information as possible in the low-dimensional space, and thus it achieves the best recognition performance compared to the other methods.