Research On Human Action Recognition Method Based On Deep Learning

Human action recognition has been applied in many fields and is a hot research direction in computer vision science.Action has diversity and complexity in the spatial and temporal dimensions,and action recognition is the key to some other fields,so the research on action recognition still has broad prospects.The combination of deep learning technology and action recognition allows researchers to design flexible and diverse recognition methods.Based on deep learning and skeletal data of action sequences,we first propose an action recognition method of enhanced view-independent action feature representation and combined temporal attention model.Combining the cumulative Euclidean distance of nonhip joints enhances the action features and highlights the joints that are more closely related to the action.Then,a combined temporal attention model is used to extract features and classify actions.Experimental results on two benchmark multi-view datasets show that this enhanced viewindependent action feature representation can improve recognition accuracies and better robustness compared to skeletal data,and the entire method can achieve good recognition accuracies.To further improve the robustness and reduce confusion of actions,we propose an action recognition method based on frame interpolation and combined self-attention model,retaining the combined temporal attention model architecture.The frame interpolation operation expands the information of the action sequence in the temporal dimension and improves the robustness and recognition accuracies.The combined self-attention model makes full use of the advantages of transfer CNN in image recognition and mitigating overfitting.The self-attention mechanism automatically focuses on significant distinguishable information.Compared with the former method,this model can improve recognition performance.We also explore real-time action recognition.A real-time action recognition method based on keyframe selection was proposed.Although it is not ideal in recognition accuracy,it is significantly effective in reducing the amount of frame input data,and its real-time evaluation indicator is far better than the first two methods,which has improved value in future work.