Identification And Localization Of Human Backed Acupoints Based On Deep Learning

With the aging of my country’s population and the increase of life pressure,more and more people use traditional Chinese medicine moxibustion rehabilitation physiotherapy,in view of the problems of high cost of manual acupoint search,strong professionalism,long moxibustion time and repeated labor in the traditional moxibustion treatment process.In this paper,we first obtain high-accuracy and highprecision acupoint coordinates by using HRNet and PFLD acupoint recognition deep learning network,Then,through camera calibration and hand-eye calibration,the pixel coordinates of the acupuncture points on the back of the human body are converted into three-dimensional coordinates in the coordinate system of the robotic arm,finally,control the AUBO i5 robotic arm to reach the target acupoint for moxibustion.First,aiming at the problem of acupoint identification on the back of the human body,a data set on the back of the human body was constructed,in which the label set marked 43 acupuncture points.The dataset was used to train and test the HRNet and PFLD acupoint recognition networks.The average recognition rates of the networks were 96.3% and 98%,respectively,and the average pixel differences were 1.95 and1.53 pixels.The results were visualized,the experimental results verify the effectiveness of the network.Finally,the lightweight and high-accuracy PFLD acupoint recognition network is selected by comparison.Secondly,in order to obtain the three-dimensional coordinates of the acupoint points in the robot arm coordinate system,the Kinect2 camera was calibrated by Zhang Zhengyou’s calibration method to obtain the internal parameters,and the color camera and the depth camera were registered,the hand-eye relationship matrix is obtained by using the AR labeling method for hand-eye calibration,and the conversion of the pixel coordinates output by the acupoint recognition network to the three-dimensional coordinates of the robotic arm is realized.Finally,in order to realize acupoint search in the real environment,an experimental platform for acupoint identification and localization was built.The Kinect2 camera was used to collect three posture images of the back of the human model,and the acupoint coordinates were obtained through the PFLD acupoint recognition network.The accuracy of the three postures was calculated to be 95%,90.6%,and 93%,respectively.Through the AUBO i5 robotic arm,the selected 10 acupoints that meet the accuracy requirements are used to search for acupoints to verify the reliability of the robotic arm for acupoint search.