Human Action Recognition And Classification Based On Video Dynamic And Static Features And Deep Learning

Human Action recognition and classification under complex backgrounds in video have a certain degree of interference in the extraction of trajectory features of key people's movements.The static features generated by key frames and the dynamic features generated by trajectories in video clips can better describe the action of people in the video.Therefore,designing a human action recognition and classification system based on dynamic and static features and deep learning to achieve autonomous learning of the character trajectory features in the video can more comprehensively analyze the different features produced by the character movement,which has practical significance for improving the recognition and classification accuracy of character movements in different sports scenes.This thesis aims to the recognition and classification of human motion in video.Firstly,a feature extraction method based on trajectory optimization is designed to describe the movement of people in the video,and then a feature fusion method based on the dynamic and static features of video is used to design the movement of key people.Trajectory and background information are represented for dynamic features and static features,respectively.Finally,an improved two-stream C3 D action recognition method is designed to complete the recognition and classification of video human motion.The innovation of this thesis is mainly reflected in the following three aspects:(1)Using the human detection box detected by Faster R-CNN to extract the motion trajectories of key people,and then calculate the cosine similarity between the trajectories of the key people to remove duplicate trajectories,using FV coding to encode the low-level trajectory features,the SVM outputs the classification results.The human detection box experiment and motion recognition experiment is carried out on the KTH dataset and UCF 101 dataset,respectively.The experimental results show that Faster R-CNN can effectively detect the human detection frame.In a simple environment,it can effectively use the optimized trajectory and can more clearly characterize the trajectory characteristics of the person's movement,and the recognition accuracy and efficiency are batter than the original i DT method.(2)Using Alex Net-based convolutional neural network to extract static features from video frames,constructing action tubes to extract dynamic features from the video by trajectory,and using Cholesky changes to feature fusion,and finally compensating by GRU network.Feature fusion experiments were performed on UCF101 dataset and Hollywood2 dataset.The experimental results show that the video has the highest recognition accuracy when the contribution ratio of static features and dynamic features is 8: 2.At the same time,because using the action tubes to cluster the trajectories in each significant moving areas,the trajectory information in the video can be better used.(3)Using 3D Spetial Pyramid pooling layer to replace the last largest pooling layer in the original C3 D network,so that variable-scale video input can be available.Add a C3 D network to extract optical flow features in the video to add a C3 D network to extract optical flow features in the video to extract the feature information.Finally,experiments are performed on the UCF101 and HMDB51 datasets through early fusion and late fusion.The experimental results show that the C3 D network using 3D pyramid pooling can accept variable-scale video input,and the recognition accuracy is also slightly improved;Early fusion is better than late fusion in Dual flow improves the feature fusion of C3D networks.