Research On Action Recognition Method Based On Improved Dense Trajectories

Human action recognition technology is a process of analyzing and processing a video containing human action through a computer,then recognizing human action in the video,which is widely used in various fields,such as intelligent video surveillance,human-machine Interaction,sports analysis,content-based video retrieval,etc.Human action recognition have become one of the hot topics in the field of computer vision and image.At present,although there are a large number of action recognition methods,due to the interaction of human action,the complexity of the background,the variability of illumination,and camera shake,the slower processing time and the lower recognition accuracy of human action recognition are caused.It has not been well resolved,so human action recognition is still a challenging subject.The traditional human action recognition method based on dense trajectories is to randomly sample on dense grids,which leads to the acquired feature dimension is too high,the calculation amount is large,and the extracted trajectories contains more background redundancy information.The real-time performance is poor.In order to improve the real-time and recognition accuracy,this paper studies and improves the action recognition method based on dense trajectories.The specific research work is as follows:(1)Aiming at the long time of feature points matching processing in the Speeded-Up Robust Features(SURF)algorithm for camera motion estimation,an improved SURF algorithm based on dynamic Gaussian pyramid is proposed.Mainly to improve the traditional SURF algorithm,including two aspects: one is to construct a dynamic Gaussian pyramid;the other is to introduce the brightness center algorithm to determine the main direction of feature points and the Rotated Binary Robust Independent Elementary Feature(rBRIEF)generate a feature descriptor.The experimental results show that compared with the traditional SURF algorithm,the improved method has shorter matching time and higher matching accuracy of feature points.(2)Aiming at the problem of low recognition rate and long time caused by camera motion problem in the process of feature extraction,a feature extraction algorithm based on improved SURF is proposed.In this paper,after the dense sampling of the video,the camera motion estimation with improved SURF is introduced,and the optimized features are extracted to further eliminate the interference caused by the background redundancy information,thereby improving the real-time and accuracy of feature extraction.(3)Finally,the fusion of features and the recognition algorithm of actions are studied,mainly for the single feature extracted and the low feature recognition rate of direct combination.A feature weighted fusion method is proposed to express the features and identify the action.The analysis and evaluation of the experimental results show that the method proposed in this paper has certain advantages.