Research For Human Behavior Recognition Based On Video

With the increasing maturity of Internet technologies,especially the promotion of mobile Internet applications and the popularization of video surveillance,human-computer intelligent interaction,video retrieval,and medical fields,video has become an indispensable media form in people's daily production and life.However,human behavior in real scenes is uneven.It is still a challenging problem to accurately identify and analyze human behavior in real scenes,which is influenced by objective factors such as different perspectives,chaotic scenes and blurred photography,and subjective factors such as high complexity and variability of human behavior.Therefore,how to effectively improve the accuracy of behavior recognition has become a hot research direction in this field.Because effective features can efficiently characterize the current behavior of the human behavior,it is very important to develop an advanced behavior recognition algorithm.Aiming at the research direction of human behavior recognition based on video,the thesis carries out relevant research and Exploration on traditional methods and deep learning methods.The main research work of the thesis is as follows:(1)Improved Dense Trajectories for Action Recognition based on Random Projection and GMM-FV Hybrid ModelsThis paper introduces an advanced vector method that combines Fisher's vector with random projection when extracting features using the Improved Dense Trajectory Method.The method firstly reduces the number of tracks and removes redundant moving information by moving the boundary sampling method,and then uses the random projection algorithm to project the high-dimensional trajectory descriptor into the low-dimensional subspace to realize the dimensionality reduction of the feature trajectory.Finally,the random projection is used to perform a second dimensionality reduction on the GMM-FV hybrid models to reduce computational complexity.The experimental results show that compared with the traditional identification methods,this method not only reduces the complexity of the computer,but also improves the accuracy of behavior recognition.(2)Multi depth feature fusion based on of C3 D network for video human behavior recognitionExtracting efficient and comprehensive behavioral characteristics is the key to behavior recognition.Traditional single behavioral characteristics often involve only part of the behavioral data,and other information that can characterize behavior is discarded,which resulting in a certain impact on the accuracy of behavioral recognition.Therefore,this thesis proposes a video behavior recognition method based on C3 D network for multi-depth feature fusion.C3 D network is used as feature extractor to extract the depth features of optical flow and RGB respectively.Then feature fusion is carried out using feature weighted fusion method to generate the final feature vector.Finally,libSVM classifier is used to train and classify.The experimental results show that the improved method is more effective than the traditional 3D convolution network to improve the recognition rate of behavior recognition.(3)Video behavior recognition based on LRCN network structureTo further investigate video behavior recognition,this paper evaluates video behavior recognition based on long-acting recursive convolutional network.Which combines the convolutional layer not only to process variable input sequence,but also to support variable output,and can generate complete statement description,so as to capture information on time seriesThe experimental results show that although LRCN model has advantages in dealing with long input sequences,the final classification accuracy is not as good as C3 D network.