Research On Skeleton Based Lightweight Action Recognition Model

With the advancement of artificial intelligence technology and the upgrade of robot software and hardware,coupled with the impact of the COVID-19 and the catalysis of the home economy,home service robots have gradually penetrated into human life.Intelligent service robots need extensive human-computer interaction and intelligent monitoring in their daily work scenarios,so human action recognition has become the basis of inclusive home service robots.Considering the limited image processing capability of current robot onboard processors,lightweight action recognition model has important research and application value.Through the research on the skeleton-based lightweight action recognition model,this paper designs optimization strategies in feature representation,feature extraction,model lightweight,etc.This paper verifies and analyzes the optimization strategies on public datasets.The main work of this paper is as follows.First of all,according to the characteristics that human motion information is concentrated and skeleton end joints contain more key information,this paper proposes a lightweight skeleton action recognition model based on attention mechanism.This model introduces the idea of dual attention,which combines the attention of spatial domain and channel domain.On the one hand,it improves the feature expression of key spatial locations,enhances specific target regions of interest,and at the same time weakens irrelevant background regions;on the other hand,it enhances the importance of key feature channels and enhances the capture of key spatial features by graph convolutional networks.So that a shallower network structure can be used to improve the accuracy of action recognition and achieve the goal of lightweight network.Secondly,aiming at the problem that the joint features of skeleton sequences carry limited information,a lightweight skeleton action recognition model based on multimodal feature fusion is proposed.The model not only extracts more informative and discriminative skeleton features,but also extracts the angular feature that distinguishes similar motion trajectories.At the same time,in order to avoid that the fixed topology structure is difficult to describe various action samples,a graph structure that can be dynamically adjusted according to the data samples is designed,which further improves the flexibility and accuracy of the action recognition model.Finally,on the NTU RGB+D 60 dataset and NTU RGB+D 120 dataset,this paper conducts experimental verification and result analysis on the lightweight skeleton action recognition model based on attention mechanism and the multimodal feature fusion.The experimental results demonstrate the effectiveness of the dual attention mechanism 、multimodal feature fusion and adaptive graph convolution layer,which can improve the accuracy of the recognition model while maintaining the lightweight characteristics of the model.