Research On Human Action Recognition Method Based On Deep Learning

Action recognition has been widely used in the fields of human-computer interaction,virtual reality,video surveillance,and video retrieval and analysis,which has attracted more and more researchers' attention.Action recognition has important academic research value and strong practical value,and is the research hotspot and difficulty of computer vision,pattern recognition,artificial intelligence and other fields.The exist problems of action recognition are as follow: There are large intra-class variations in the same action class,which may be caused by background clutter,viewpoint change,and various motion speeds and styles;Moreover,some actions include similar motion patterns,which make different classes of actions have less interclass variation and further cause confusion.Meanwhile,feature redundancy caused by high-dimensional video data,camera motion and low resolution of video further increase the difficulty to extract effective features and design robust recognition method.How to extract efficient features from videos and construct a more effective action recognition framework is a key issue to be solved urgently.In this thesis,the existing methods of action recognition are analyzed and summarized,and the following works have been down:Firstly,the common action recognition methods are analyzed and summarized.As for the problem that traditional descriptors do not consider the joint statistical characteristics between features,the time derivatives of image gradient,optical flow and motion boundary are taken as low-level motion features on the basis of dense trajectory.And the covariance matrix between low-level features is calculated to construct Trajectory Based Covariance Matrix(TBCM)descriptor,which takes full account of the joint statistical characteristics between features and further improves the descriptive power for behavior subject in complex environments.Secondly,a discriminant nonlinear feature fusion method is proposed.The category structure information is introduced into the objective function of the Kernel Canonical Correlation Analysis(KCCA)to construct a new feature fusion method.This method maximizes the nonlinear correlation between global and local features,reduces the intra-class variation,increases the inter-class variation,and thus further enhances the discriminant ability of features.Thirdly,deep 3D convolution descriptor is constructed.In this thesis,the feature vectors of any a layer in Convolutional 3D network(C3D)are extracted.The fully connected layer-6 and layer-7 feature vectors are respectively extracted and concatenated to be taken as global feature,and the pooling layer-4 and pooling layer-5 feature vectors are respectively extracted and concatenated to be served as local feature.Through the proposed discriminant nonlinear feature fusion method,global and local features are fused to obtain a more complete deep 3D convolution descriptor.The proposed methods are compared with existing methods and verified on UCF-Sports and You Tube datasets.The experimental results demonstrate the effectiveness of our methods.Finally,the main works of thesis are summarized,and the future research direction is also given.