A Human Daily Action Recognition Method For Controlling A Home Care Robot

Home care robots are very helpful in alleviating the depression of the elderly and other people living alone,reducing psychological pressure,improving the users’ initiative in life,and other social assistance.In the research field of home care robots,how to make the robot quickly and accurately identify the user’s behavior and actions and how to maintain the availability and security of the robot control system has become one of the important research topics in this field.Therefore,this paper focuses on the key technologies of human behavior and action recognition and the human-robot interaction process based on quadruped robots.This paper analyzes the problem that the data set of the training network model in the current popular Open Pose human posture estimation method cannot be identified or misidentified due to incorrect labeling,and proposes an Open Pose human posture estimation method based on label correction,which is applied to the development of robot control system in the scene of home care.In this method,the original method is firstly used to generate the model’s first label,and then the model’s first label prediction results are used to modify the labels of each pixel in the image during the second repetition training.The online experimental results on public data sets show that the mean average precision of the improved human posture estimation method increased by 3.4%.The offline experimental results in the home scene show that this method can still accurately infer the two-dimensional image position of the user’s whole body nodes when part of the user’s joints are covered by objects,and the positioning and tracking of the human body are not affected by occlusion.In order to avoid redundant feature extraction,part of the nodes were selected as static features through human kinematics analysis,and angle features,length features and motion speed features were calculated as dynamic features according to the data information provided by the nodes.Aiming at the problem that the accuracy of behavior recognition is low and the real time performance is difficult to be guaranteed,a two-level classification method based on the fusion of support vector machine and convolutional neural network is proposed in this paper.The extracted multi-dimensional human body node data and the deep-level features of skeleton graph are classified respectively,so as to complete the recognition of human daily behavior and actions.The experimental results show that compared with the single method,the average time per frame of the proposed method is reduced by 0.174 s at most,and the accuracy rate is increased by 12.7%.This method has good performance in accuracy and real-time performance,which can meet the interaction requirements of the control system of the home care robot.The size of the sliding window has been modified according to the duration of a complete motion to accommodate different duration of continuous motion detection.In order to verify the actual interaction effect of the control system of the quadruped home care robot based on the recognition of human daily behaviors and movements,the experimental verification research is carried out after the system integration.In the actual home scene,comprehensive experiments of control system’s availability performance are performed.The experimental results show that the control system of the home care robot can capture multiple types of human daily actions in real time,and generate the service instructions for the home care robot synchronously according to the detection results to realize the intelligent service of the robot.The recognition success rate of the system is 93%,and the execution success rate is 90%.Finally,the user’s experience of the system is evaluated and the user psychophysical experiments are conducted.The participants of the experiment have high evaluation on the reliability and effectiveness of the system.