Research Of Human Pose Estimation Method Based On Convolutional Neural Network

In the existing mature robot technology and products,the robot in the human-machine collaborative operation space does not estimate human upper limb pose that cooperates with it,and thus cannot guarantee human's safety.The robot in the human-machine activity space does not estimate multiple human dynamic pose,and thus cannot predict the trajectory of the human movement,i.e.,it cannot avoid collision with human.This thesis investigates the estimation method of human pose in the human-machine collaborative robot system.We have estimated the single body upper limb pose in the human-machine collaborative operation space and the multiple body pose in the human-machine activity space,and have solved issues concerning human security in two spaces.To be specific,our study can be summarized as follows :For the single-limb pose estimation problem in the human-machine cooperative operation space,based on the convolutional neural network,this thesis adopts an estimation methods of single body upper limb pose based on end-to-end mode.Based on the stacked hourglass network model,the detection models of the single body upper limb's skeleton keypoint in the single upper limb pose estimation method has been designed.The calculation method of the position coordinates of the hourglass module and the skeleton keypoint has been improved.For the multi-person pose estimation problem in the humanmachine activity space,this thesis adopts an estimation method framework of multi-person pose with the multi-person pose estimation using pose residual network.Based on the YOLOv3 target detection model,the human detection model in the framework has been designed,and the non-maximum value suppression algorithm and the size of the a priori frame in the detection model have been improved.At the same time,the human skeleton keypoint detection model in the multi-person pose estimation method using pose residual network is adopted,and the pose residual network model has been adopted as the grouping model of the human body's skeleton keypoint in the framework.For the integration of human pose estimation system,this thesis makes use of the communication mechanism with Publisher/Subscriber and Client/Service.Moreover in the robot operating system,integrates the estimation method of human body pose into the robot operating system,so that the system could be directly used as a subsystem of the human-machine collaborative robot.By conducting test experiments on single and multi-person image samples,it is verified that the method used in this thesis can estimate the single-limb pose and multi-person pose more accurately and real-time,and has higher practical application value.