Dance Action Recognition Based On Pose

The study of human action recognition can be summarized as three levels according to the action content from simple to complex:mobile,action and motion.At present,there are two problems in the study of video human motion.First,the repetition information of the majority of frames in the video or the low correlation with the recognition,not only increases the computational complexity,but negative of recognition accuracy.Second,in the process of feature selection and understanding,the main methods include characterization and integration based on attitude characteristics.But the stage of the dance and the clothes changing,and in the performance process prone to occlusion and self-occlusion problem,so the characterization of information fusion in most cases can not accurately and completely express the body motion information,and based on posture characteristics are easy to ignore the human static information.In our paper will have realized the application of pose-based action recognition technology in dance action,and have finished the recognition of northeast Yangge dance movement.Firstly,the motion segmentation algorithm based on pose is proposed to reduce the repetitive and irrelevant actions in video.According to the posture information of the human body,the trend curve of the continuous posture of the human body is formed,and the extreme position of curve is determined as the important action position in the curve.By solving the curve cubic spline interpolation function elimate the continuous extreme value.Use of the frame number and the continuous extreme difference to determine the position of the curve that is the minimum position of the video frame.The DTW method is used to calculate similarity between the action sequences in the same video,and it is judged whether the different action sequences belong to the same or similar action.Then,it is proposed to use the pose-based feature description to fully represent the human action information.The human body region in the frame is obtained by using the human posture,and the 3D-SIFT and optical flow characteristics areextracted for each region.Where SIFT represents the human body static information,the time dimension of join can obtain continuous human static information,easier to represent the human body dynamic information with the use of optical flow characteristics.At the same time SIFT illumination invariant feature can compensate for the light flow characteristics of light sensitivity,while the optical flow as a SIFT key point selection based on the ability to obtain a more stable key point.Finally,the feature of the extracted feature is reduced by normalization,and the eigenvector of each kind of feature is obtained.Then,the feature vector is fused as input of classifier to realize recognition of dance action.In the northeastern yangko dance data set,the proposed method based on pose action sequence segmentation can obtain high precision and recall rate.The combination of static information and dynamic information is accurate and complete to express the information of human action in video accuracy of recognition.