Group Sparse Coding And Dictionary Learning Based Human Action Recognition

With the high-speed development of the modern society, people's lives are filled with a lot of video data. And human action recognition is not only a important topic in computer vision, but also has very broad application prospects in areas such as in video surveillance,human-machine interaction, medical assistance and so on. This dissertation aims to recognize human actions.Due to the video setting, video angle change, the video light changes and the impact of the presence of occlusion, so that the computer can not be good for the video to identify human behavior, although human action recognition has been a great development. The research of human action recognition consists of three parts: human action representation in videos, the learning model of human behavior and the classification model of the human behavior. Human action representation is the basis of the whole process, which directly affects the final recognition result; the classification accuracy is affected occlusion problems and so on, so the key is how to learn a discrimination model. According to the difference among the used sample,, the methods of human activity recognition can be summarized as three classes: supervised learning method, unsupervised learning methods and semi-supervised learning methods. Since the supervised methods have achieved better recognition results than unsupervised methods. And taking into account real life filled with a lot of unlabeled samples, while the cost to obtain the sample is enormous. So this thesis is mainly to study how to effectively use the semi-supervised method to feature encoding.The main contributions of this paper are as follows:1. We propose a semi-supervised dictionary learning method based on similarity weights for human action recognition. By constructing the similarity weights between the video samples and the encoding dictionary, the method can introduce the information of unlabeled samples, so that the semi-supervised learning method and the behavior recognition are combined to learn the discriminative dictionary.2. We propose a group sparse coding method based on local 2,1-norm for human action recognition. 2,1-Norm is a concept of row sparse of the coding matrix, it can make a dictionary atom or to participate in encoding of all local features, or not to participate in encoding of any local feature. In fact, the local features of a video have its local similarities.We propose a local 2,1-norm, which not only considers the overall sparsity of the video,but also takes into account the local information in the video. The proposed method uses a group sparse representation model, using the local 2,1-norm, for the video to feature encoding.3. We propose a semi-supervised dictionary learning method to improve the performance on human action recognition. The difference between supervised dictionary dictionary learning and unsupervised dictionary learning is whether the label samples are used.Semi-supervised dictionary learning not only to use the label samples, but also to use unlabeled samples. In this paper, the labeled sample and similarity constraint of the dictionary are used to improve the discrimination of the coding dictionary, which can be obtained by using a large number of unlabeled samples, then it can get the better effect than the supervised and unsupervised learning method.