Human Motion Recognition Based On Depth Information Coding

With the rapid development of science and technology and the improvement of people's living standard,people are becoming more and more concerned about the intelligence of life.Human motion recognition is widely used in intelligent monitoring,human-computer interaction,motion analysis and so on.Human motion recognition is a challenging scientific research task in the field of computer vision,which involves many disciplines such as image processing,neural network,machine learning and so on.However,due to the complexity of the background,the camera movement,the intensity of light changes,the different clothes,as well as the existence of occlusion problems,resulting in human motion recognition is difficult to achieve.At present,the field of human motion recognition is still in the research stage,and the database background is relatively single,the number of video is limited,most of the videos are only one subject,the body action recognition distance practical application there are many ways to go.In view of the current human action recognition is proposed scheme is mostly fixed input,large amount of calculation,high time complexity and the lack of an effective use of sports information and information of time and space,this paper deeply studies the human action recognition based on the deep learning,which combines optical flow,Spatial Pyramid Pool,feature coding and so on.The paper has done the following three parts:1.A method of human motion recognition based on optical flow and SPP-Conv Nets is proposed.The video and optical flow as input to train the SPP-Conv Nets network,then by calculating the optical flow intensity box to the target areas,finally using the SPP-Conv Nets model is trained to test out the video and optical flow frame of the target areas.This method makes the input image for any size,the image distortion caused by interception and stretching is reduced.The static information and motion information are fused to improve the recognition rate and the robustness of the algorithm.2.A method for human action recognition based on Two-stream network and local deep coding is proposed.First of all,the CNN feature map of the video and optical flow are extracted by Two-stream network.Secondly,the local depth feature descriptors can be obtained by using local feature encoding.Local feature encoding preserves the spatial information layer includes volume and the increase of the number of descriptors can improve the accuracy of human action recognition.In order to train K-means,we first de-correlate the local depth feature descriptors and reduce its dimension.The video and optical flow can be represented with VLAD vector.Finally,combining the video and the corresponding optical flow VLAD vector as the input data of the SVM classifier.3.Based on spatial linear pool,the linear feature descriptors are obtained by using the method of spatial linear pool to encode the CNN feature map.In order to train K-means,we first de-correlate the local depth feature descriptors and reduce the dimension.Each video and optical flow are represented by VLAD vector respectively.Finally,combining the VLAD vector of the video and the corresponding optical flow,and using SVM to classify and identify.The spatial linear pool encoding into the pool of linear trajectory information based on static information and motion information,more in-depth digging trajectory information among consecutive frames in the video to improve human action recognition accuracy,and the algorithm is simple,small amount of features.