Research On Key Techniques Of Human Action Analysis In Videos

In the past thirty years, computer vision has been developing rapidly. As currently one of the most active research topics in the domain of computer vision, the analysis of human behavior has been widely used in the field of computer vision, such as intelligent monitoring, human-computer interaction, virtual reality and content-based video retrieval and interpretation. With the continuous expansion of its increasing demand of application, the research of improving its robustness to complex environmental factors and the validity of the algorithm of human target feature extraction and description becomes the research difficulties and key issues.Firstly, in the aspect of object feature extraction and description of moving human motion target, a method of improving the performance of dense point trajectory is proposed. Based on the feature of dense point trajectory and the consideration of the defects of camera motion, when extracting the SURF and optical flow field, we use the random sample consensus algorithm to match the feature points to eliminate the influence of the camera movement, which can improve the recognition results.Secondly, in the aspect of human action recognition algorithm, a human action recognition algorithm based on dictionary pairs learning method was researched, through the method of the introduction of auxiliary domains, which obtains a dictionary pair, effectively expanded intra class diversity of the training sets. This paper uses improved dense trajectory to describe the low-level features. Through the cross-domain dictionary pairs learning, we acquire corresponding sparse representations of the actions. The algorithm for dictionary learning and classification is a combined framework. We use the reconstruction error of the dictionary learning algorithm to classify.Finally, experiments are carried out under the conditions of MATLAB simulation. The two sets of experiments are carried out for human behavior feature extraction and human behavior recognition algorithm. As for the experiment of human action feature extraction and description, through the comparison of SIFT point trajectory and KLT trajectory, it is proved that our trajectory do well in the fast-change and irregular movement patterns. As for Human action recognition algorithm, we selected the UCF YouTube dataset as the original training set, HMDB51 data set as an auxiliary domain, seven corresponding actions are selected from the two action data set, these actions are cycling, diving, golf, jumping, batting, riding horse, shooting. According to the process of the algorithm, feature extraction and recognition, by the comparison with related algorithms, the recognition results of our frame are significantly improved. It is proved that the effectiveness of the algorithm.