Font Size: a A A

Research On Low-rank Presentation And Recognition Of Human Actions In Video Sequences

Posted on:2016-05-04Degree:DoctorType:Dissertation
Country:ChinaCandidate:S J HuangFull Text:PDF
GTID:1108330503452398Subject:Instrument Science and Technology
Abstract/Summary:PDF Full Text Request
In computer vision, human action recognition has become one of the most active research topics with great theoretical research value and many important applications, such as intelligent video surveillance, video retrieval, intelligent monitoring, and human-computer interaction. Recently, many approaches for action recognition have been proposed, which explain and deal with action recognition from different perspectives and have achieved some successes. However, due to the complexity of human action and its environment, there are still many problems worthy of further study. Human action recognition was usually regarded as a classification problem, which consists of feature extraction and representation and action classification. In this paper, we take the video sequences containing human action as our research object, and analyze the characteristics and shortcomings of existing action recognition approaches. To overcome these shortcomings, we deeply study the extraction and representation of action feature and corresponding action recognition approach, and present some novel solving ideas and methods. In this paper, we first propose action low-rank feature. Afterwards, we present accumulated edge distribution histogram to represent the low-rank feature. Then, we present two improved approaches to enhance its ability of resisting background interference and capturing temporal information. We also study deeply the corresponding action recognition methods, and propose novel models of learning discriminative parts and learning dictionary. Finally, the experimental results on three benchmarks demonstrate the effectiveness of the proposed approaches.The main contributions and innovation points of this paper are summarized as follows:① This paper presents action low-rank feature, and determine the feasible parameter and its formula for extracting action low-rank feature. Traditional action features usually require some intermediate processing steps such as actor segmentation, body tracking, and interest point detection. However, there are still many unresolved difficulties in these intermediate processing steps, the errors of which will reduce the final recognition performance. Compared with traditional action features, the action low-rank feature has more concise extraction manner, and can effectively avoid the intermediate processing steps above. However, it is difficult for traditional regularization parameter to capture effectively the action information included in video sequence. Thus, we conduct extensive experiments and determine the feasible parameter and its formula for extracting action low-rank feature. Experimental results demonstrate that the effectiveness of action feature extraction of our presented regularization parameter is far better than that of traditional regularization parameter.② Accumulated edge distribution histogram is proposed to represent the action low-rank feature. Due to the characteristics of action low-rank feature, traditional representation methods cannot effectively descript the action information in the action low-rank feature. The edge information of action low-rank feature can effectively overcome the influence of grey information of actor’s clothing, and can be generated by the human motion in video sequence. Thus, we first extract the edge information of action low-rank feature. Then we perform a statistical analysis on the distribution of edge information, and finally form accumulated edge distribution histogram representation. The experimental results on three benchmarks demonstrate that the presented accumulated edge distribution histogram was more suitable for representing the action low-rank feature than traditional methods.③ An action recognition approach is proposed based on discriminative part learned from action low-rank feature. There was a strong complementarity between discriminative part learning and action low-rank feature. Learning discriminative parts from action low-rank feature can effectively enhance the ability of resisting background interference of action low-rank feature, and greatly overcome the problem of background memory suffered by traditional part learning models. However, the traditional part learning models usually learn identical part number for all action categories, which neglects the differences of recognition difficulty of different action categories. Thus, we propose a novel learning model of discriminative part with flexible number, and this model can learn discriminative parts with flexible number for each action category. In this model, we define a new similarity constraint to promote the generation of discriminative part detectors, and employ group sparse regularizer to automatically reserve the strong discriminative part detectors for each action category. Experimental results demonstrate the effectiveness of our proposed part learning model and the better performance of our action recognition approach.④ An action recognition approach is proposed based on temporal action low-rank feature and dictionary learning. To capture the time information in video sequence, we first propose temporal action low-rank feature. Specifically, we first split an action sequence into multiple overlapping subsequences according to an overlapping ratio. Then, we extract the low-rank feature of each subsequence, and concatenate all low-rank features according to their time order to form temporal action low-rank feature. Dictionary learning model can be employed to classify the temporal action low-rank features. However, traditional dictionary learning models neglect the similarity constraint among the coding coefficients of different samples, and have bad performances with non-linearly data. To overcome these shortcomings, we propose a novel similarity constrained discriminative kernel dictionary learning model. In this model, a new similarity constraint is defined to force the coding coefficients of different samples, which is conducive to training better classifier. Besides, a nuclear mapping method is added to the model to enhance its ability in processing non-linearly data. Experimental results demonstrate the effectiveness of our proposed dictionary learning model and the corresponding action recognition approach.
Keywords/Search Tags:Action recognition, Action low-rank feature, Discriminative part learning, Temporal low-rank feature, Dictionary learning
PDF Full Text Request
Related items