A Research Of Human Posture Estimation And Manipulator Motion Planning Based On Neural Network

Human pose estimation based on computer vision mainly studies the processing of the original images acquired by the camera.After these images are preprocessed,it detects the main joints or key parts of the human body in the image,and outputs the relative positions of the various joints about the human body.Finally,it learns and understands the human body's behaviors,and put it into practical application scenarios such as pedestrian detection and movie special effects.Depth cameras,such as Kinect,can obtain depth information of human behaviors,which is convenient for computers to recognize and process these human postures and perform human-computer interaction.This interaction method can achieve relatively stable results,but it is extremely dependent on hardware equipment,and the application scenarios are subject to limit.In the current social environment,ordinary cameras are widely used.Therefore,studying the stability of monocular camera pose estimation and application has good social and economic benefits.Based on the estimation of the human body posture,this subject completed the study of the robot arm highly imitating the human arm.The main contents of this thesis are as follows:1.The technology of human pose estimation based on computer vision is studied.Firstly,the basic structure of MobileNetV2 is analyzed,the activation function in the network is replaced,and the advantages of ShuffleNetV2 network structure are borrowed,the network structure of MobileNetV2 is improved,and the recognition rate and efficiency of the network are improved.Then,Then,a pose estimation network model based on improved MobileNetV2 is constructed.Finally,the data samples are used to complete the network training,and the recognition results are compared and analyzed.2.The kinematics of the manipulator is studied,and the kinematics model of the manipulator is established.Firstly,UR3 is used as modeling target,the mathematical model is established through using the standard D-H parameter method by Matlab.Then,the inverse kinematics of the robotic arm is solved according to the geometric transformation and the space vector method.Finally,in the joint space,the trajectory planning of the robotic arm is completed and the simulation experiment of the trajectory planning is carried out.3.The manipulator imitating the human arm experiment is carried.Firstly,a 3D human arm information recovery method based on image ranging and spatial geometry is proposed.Then,the human-machine mapping model according to the space vector method is build.Finally,the imitation experiment is completed,and the similarity evaluation method of imitation results is given.After conducting multiple sets of experimental tests,the results show that the improved MobileNetV2 based on human body pose estimation,recognition efficiency increased by 40%,recognition success rate reached 92%,and it's stable.The robotic arm can completely follow the movement of the human arm,and the similarity reaches 90%,which proves the feasibility of the technical solution of this thesis.