Design Of Human Action Recognition System For Intelligent Nursing Robot

In the field of intelligent nursing robot,how to quickly and accurately recognize the actions of the nursing object has become a hotspot in this field.Accurate recognition of behavior is not only the premise for nursing robot to realize intelligentized nursing work,but also can improve the situational awareness of robot.Thence,action recognition technology is an important part in realizing intelligence of nursing robot.Action recognition methods based on deep learning are equipped with the advantages of simple modeling process and easy training.It has gradually become the development trend of action recognition.However,above methods are limited by accuracy of action recognition,generalization and convergence speed.Based on mentioned research background,this paper combines graph convolutional neural network and attention mechanism to carry out research and proposes a new human action recognition system for intelligent nursing robot.(1)With the purpose of explore the feasibility of applying graph convolutional neural network to human action recognition,the researches on the characteristics of human action are carried out.We also analyze the traditional convolutional neural network and graph convolutional neural network;Then,by combining multi-feature learning method and feature fusion strategy,a novel action recognition model based on two-stream graph convolutional neural network is proposed to realize feature extraction of skeleton data;Finally,in order to verify the effectiveness of the model,action recognition experiments are conducted on NTURGB+D and UCF101 datasets.(2)With the aim of solving the following problem in two-stream graph convolutional neural network: the model lacks the ability to model the temporal domain for the global context temporal information;non-ideal recognition accuracy rate and non-enough generalization performance,following researches are conducted.Firstly,temporal action graph and adjacent matrix based on N-order fixed-time structure are constructed with graph representation theory and the feature representation method of the skeleton sequence;Secondly,through combining temporal adaptive graph convolutional structure with spatial adaptive graph convolutional structure,the spatial-temporal adaptive graph convolutional network is proposed in this paper.Temporal action graph is used as the input data of it;Finally,combined with the spatial adaptive graph convolution structure,the spatio-temporal adaptive graph convolution neural network is proposed.Action recognition researches are conducted on the NTU-RGB+D and SBU with the aim of validating advantages of model in capability of modeling global context temporal information,classification accuracy and generalization ability.(3)In order to extract spatio-temporal information and retain the correlation between spatial information and time information,and realize attention on specific joints for important action information,following researches are conducted.Firstly,based on the graph concolution kernel which can handle variable-length neighbor nodes in graph,a 3D graph convolution method is proposed where 3D sampling space of 3D convolution is introduced to improve 2D graph convolution kernel to 3D graph convolution kernel with 3D sampling space;Secondly,an attention enhanced structure is designed to enhance attention to specific joints and focus important action information,an attention enhanced structure is designed;Then,through combining 3D graph convolution with attention enhanced structure,action recognition model based on 3D graph convolution and attention enhanced is proposed;Finally,the researches are carried on NTU-RGB+D and MSR Action 3D skeleton action datasets.(4)In order to meet the needs of nursing robot for the perception and action recognition of nursing objects,a human action recognition system is designed and constructed.Firstly,we analyze the actions in home nursing scene and use the Kinect v2 camera to collect the action samples to construct an action recognition dataset under intelligent care environment for finetuning the action recognition model;Secondly,to verify the action recognition system for the nursing robot,the feasibility of the system is analyzed,and the hardware system and software development environment of the system is selected.Then,the action recognition system is built and the relevant experimental environment required by the system is also configured;Finally,under the home nursing scenario,examinations of action recognition system and action recognition are carried out to verify the feasibility of the system modules.