Research On Technology Of Skeleton-based Action Recognition

In computer vision,skeleton-based action recognition is one of the most popular research topic in recent years.It has great application scenarios in many aspects such as intelligent security and human-computer interaction.In addition,human skeleton information data is lighter than RGB video data,which makes it possible to use smaller algorithm models to gain better recognition performance.The thesis first investigates the current research status of this task at home and abroad,and then analyzes several ways to construct the dependency relation between skeleton joints in classic graph convolutional networks and their pros and cons.Based on this,the thesis proposes the Joints-dependencice Inference Network(JIN)and the Context-encoidng Network(Ce N).These two networks can automatically infer the adjacency matrix that is most suitable for the current skeletal sequence according to the spatiotemporal varies of the skeletal sequence.Among them,JIN infers the dependency relation between any two skeleton joints according to the spatiotemporal varies of the skeletal sequence to construct a symmetrical adjacency matrix for each skeletal sequence,and Ce N constructs asymmetric adjacencies for each skeletal sequence from the perspective of global context information.These two networks can form an end-toend network framework,JIN-SGCN and Dynamic GCN,with graph convolutional networks.In the cross-object benchmark of the NTU-RGB+D dataset,compared with the classic ST-GCN algorithm,the single model recognition accuracy of these two algorithms has been improved by 4.5% and 6.4%,respectively,which have achieved state-of-the-art.The paper also analyzes the theoretical relationship between graph convolutional networks and convolutional neural networks with channel transposition.It is obtained that when the size and number of convolution kernels are relaxed,the convolutional neural network with channel transposition is equivalent with graph convolutional network.Based on this,the paper proposes an Advanced Co-occurance Feature-learning Network(ACFN)combining grouped convolution and deep separable convolution techniques.Compared with ordinary co-occurrence feature learning networks,the recognition accuracy of ACFN is improved by 0.9%,and the parameter amount is reduced by 10.3%.The thesis also designed the entire system flow of action recognition,and introduced five modules: processing of RGB video,data processing of extracting pedestrian skeleton information and skeleton information from RGB video.Finally,the system process was tested with monitoring scene data and simulation data.