Research And Application Of Object Attitude Estimation Based On Point Clound

In recent years,with the continuous progress of artificial intelligence technology,social productivity has been greatly improved.Computer vision has become one of the most popular research fields in artificial intelligence technology.The progress of technology provides the basic foundation for many complex industrial engineering.The technology of target recognition and semantic segmentation in two-dimensional image has been more mature.At the same time,the detection and recognition technology of three-dimensional objects in the real world is also in the continuous development,using radar,3D scanner,camera and other equipment to obtain the image and depth information of the object,so as to perceive the target object on the three-dimensional plane,estimate the spatial position and posture of the object,and provide information for the manipulator grasping task.In life and industry,manipulator can replace human to complete repetitive and dangerous work,improve production efficiency and quality of human life.But in the non ideal environment,clutter,occlusion,illumination and other factors seriously affect the perception and recognition ability of the device.How to perceive objects in non ideal environment is an urgent problem.Through the depth camera to perceive the real world objects,give the computer the ability of vision,capture the image information in the real world,identify the object to be detected from the image information,according to the multi-dimensional information of the object obtained before,analyze the attitude of the object,so that the computer can have a relatively accurate estimation of the position and attitude of the object.In the robot grasping task,providing the detection information of the object can make the robot calculate the space position of the object to be grasped and estimate the corresponding attitude.This paper is divided into the following partsFirst,the LPCN rotation attitude network is constructed.Firstly,the convolution neural network is used to extract the features of the object,output the existence,coordinate information and calibration method of the object,and roughly estimate the angle information of the object.Then the object is detected again.Because the rough angle range information of the object has been obtained in the first step,the rotation angle of the object is estimated directly by convolution neural network in the second step.The mask information and rotation angle data of the object are obtained through LPCN rotation attitude network,which provides information for the subsequent 6D(6 degree of freedom)attitude estimation.Second,6D attitude estimation task.The color image information and depth image information obtained by depth camera are used as the input of dense fusion network,and the mask information output by LPCN rotation attitude estimation network is added.The RGB image data and point cloud data of target object are fused by dense pixel fusion method.The global feature extraction method is used for the point cloud data,and the lightweight attitude network output data is added to tune the estimated 6D attitude cycle.It provides 6D pose estimation information for the actual grasping task.Thirdly,the task of grabbing objects by manipulator.The depth camera Kinect V2 is used to register the color image and depth image of the object to be detected,and the point cloud information of the object is obtained.By calibrating the internal and external parameters of the camera,the 3D spatial coordinates of the obj ect are obtained.Combined with the rotation angle information and spatial position information,the 6D attitude data of the object is estimated.Combined with the reality,the object grasping experiment in no ideal environment can estimate the attitude information of the object in real time and complete the task of grasping with high real-time requirements.