Human Action Recognition In Videos Based On Local Feature Description

In recent years, video content analysis and processing has been widely concerend because of the increasing popularity of the Internet and digital devices. Duo to the wide application of the human action recognition, it has become a hot research direction. However, there are still lots of chanllenges in this research field, such as the diversity of human action, shooting background difference, movement and scaling of the camera, shooting angle difference, video noise, object occlusion and so on. They all impact the accuracy of recognition.Aimed at the problems existing in video human motion recognition, this paper proposed a human action recognition method based on discriminative parts. The method can improve the recognition accuracy through extracting video discriminant region and a middle level video description method.First of all, this paper expounds some methods of video human action recognition, including the methods based on global representation and methods based on local describtion, also then introduces the detailed process and commonly used techniques of the methods based on local representation, and analyzes advantages and shortcomings of the current commonly used methods.Secondly, this paper proposes a human action recognition method based on discriminative supervoxels. The proposed method compare one action with other actions in the regions through a pre-trained process to extract the discriminative regions of the video area as region of interest. At the same time, the proposed method describe a video based on discriminative supervoxels. The proposed method can enlarge the gap between classes and reduce the difference within classes. Moreover, it is not sensitive to the complex background, object occlusion, illumination changes. We have conducted experiments on a number of commonly used datasets. Compared with the classical methods in recent years, the experimental results show that the method achieves a high accuracy.At last, according to the visual attention mechanism of human, this paper proposed a method of human action recognition based on visual attention. At the same time, in order to solve the scaling problem in the video, we also proposed a method to represent a video by hierarchical segmentation. Experimental results show that the method can achieve a higher recognition accuracy, and is not sensitive to the video scaling.