Human Action Recognition Based On Human Key Points

Human action recognition is mainly used in many fields of computer vision,such as video surveillance,human-computer interaction,motion analysis and virtual reality,this has always been one of the hotspots of academic and industrial research.In recent years,with the development of graph convolution technology,graph convolution has been successfully applied to feature extraction tasks of non-Euclidean structural data such as social network,traffic network,protein molecular structure,etc.This also indicates that it is possible to extract the spatial features of human skeleton by using graph convolution,which is connected by human key points according to natural semantics.At the same time,cyclic neural network has made breakthroughs in natural language processing,video classification and other tasks,which also proves that cyclic neural network has great research value in extracting temporal features.Based on graph convolution and cyclic neural network,the main work is divided into the following two parts: Firstly,a new cyclic network unit is designed based on graph convolution and gated recurrent unit,which makes the extraction of spatial and temporal features alternately promote each other in the cyclic unit and further improves the recognition accuracy.Second,Designing distillation experiments,adding "knowledge" in complex models with higher recognition rate as a priori to improve the accuracy of simple models,and obtaining higher recognition accuracy on small models with faster operation speed.In order to facilitate feature extraction,this paper designs a feature mapping structure,which maps key point features to high-dimensional space to facilitate the expression of spatial and temporal features.In order to improve the robustness of the network in view,an adaptive view structure is designed,which enables the network to adapt to the key points of different views by translating and rotating them to a similar view.Attention mechanism is a common choice for action recognition tasks.In this paper,three attention mechanisms with different scopes are designed,which act on each element in data,each key point and the elements in adjacency matrix in graph convolution.Finally,the recognition accuracy of the model is further improved.Finally,the validity of the above structure is verified by experiments,and the optimal parameter setting and attention mechanism combination are found.In the actual deployment,it is better for the action recognition to be real-time with video shooting.In order to reduce the complexity of the network and maintain high recognition accuracy,distillation experiments are designed and the optimal distillation parameters are found in the comparative experiments.