Research On Human Pose Visual Recognition Algorithm Based On Model Constraints

Aiming at the requirement of high accuracy of human posture recognition by transfer service robots and the low recognition accuracy of existing human pose recognition(human joint position recognition)methods in the case of joint occlusion,this paper proposes a human pose recognition algorithm based on model constraints to solve precise extraction of human body joint space coordinates before transfering operation.First,the basic function,composition structure and control system of the transfer service robot are introduced.It is clear that the working scenario of the robot is in the home environment,and the service target is the elderly with disabled lower limbs in the family or the elderly lying in bed.It mainly implements a series of actions such as moving,holding,holding people,and placing.By analyzing the robot's navigation accuracy and the positioning accuracy of the end effector,the relationship between the accuracy of the above two and the recognition accuracy of the human posture and the safety distance is discussed.The maximum error of the human posture recognition is determined by establishing the safety distance,so the evaluation standard of the accuracy of joint recognition is determined.The choice of RGB-D camera is determined by the resolution of the camera,the maximum depth measurement distance,volume,power consumption,price,etc,and the final decision was to use Real Sense D435 as the robot's visual recognition module.For the recognition and calculation of human 3D joints,the Open Pose algorithm is first used to identify the 2D pixel coordinates of human joints in the color image obtained by the RGB-D camera,and the depth image hole repaired algorithm and median filter are used to eliminate holes and salt noise in the depth image.The 3D coordinates of human joints in the camera coordinate system are calculated by aligning the color image of the RGB-D camera with the depth image and combining the coordinate transformation.But this method is not adaptable to joint occlusion.For the recognition of occluded joints,according to the relationship between the height corresponding to the different percentiles and the size of the main joints of the human body mentioned in the "Chinese Adult Body Size" standard.The main body of the subject can be obtained by entering the height of the subject joint size.The main joint size is used as the mannequin parameter.Next,we find the direction vector between the occluded joint and the unoccluded joint directly connected to it.Using mathematical formulas to calculate the coordinates of the occluded joint from the human model parameters,the direction vector,and the coordinates of the unoccluded joint,so as to improve the recognition accuracy of human posture under the joint occlusion.When the robot performs an underarm lifting operation,it needs to extract the underarm area.According to the coordinates of the joint points identified previously,we draw the underarm ROI(Region of interest),and extract the contour of the ROI region.In this way,the closed axillary region is obtained,and the centroid is calculated as the axillary point.Finally,the joint recognition accuracy of the human pose recognition algorithm used in this paper is calculated experimentally and the experimental results are analyzed.