Research On 3D Point Cloud Based Object Size And Orientation Recognition For Mobile Robot

With the rapidly development of robot technology,mobile robot has become more and more versatile.The application of mobile robot extends to factory production,outdoor handling,military investigation,transportation and other aspects.Target object size and orientation recognition is a key technology for autonomous operation of mobile robots in an unknown environment.Object size and orientation recognition method has changed from image information based to three-dimensional point cloud information based.The main difficulty of object size and orientation recognition method based on point cloud information is how to choose a proper sensor according to the work scenes and how to construct the 3D point cloud structure of the object and extract the object information from the 3D point cloud data mixed with ground and surrounding environment information,and then accurately identify its position and size.Aiming at the demand of grasping or transporting square box-like objects in outdoor environments,using RGB-D camera single frame point cloud,an object size and orientation recognition algorithm based on Euclidean clustering of 3D point clouds and RANSAC(random sampling consistency)boundary fitting is proposed.First,the data is preprocessed by point cloud truncation,voxel filter downsampling,and outlier removal method,then ground plane point cloud is segmented.Then a fast clustering algorithm based on K-D tree is used to perform the segmentation between objects and objects,and then the 3D points of the segmented objects are projected into two-dimensional point cloud.Since the density of the edge points of the 2D point cloud is much higher than that of the internal point cloud,the RANSAC algorithm is used to fit the edge to determine the object structure.The straight line can approximate the expected mean value of the point cloud distribution of the object edge and is less affected by the point cloud noise.Experiments show that the algorithm has high accuracy,high speed,and good robustness.It can be used in mobile robots grasping or handling square objects.In indoor scenes,in order to solve the problem that a single frame point cloud sometimes has limited perspective and is not sufficient to see the overall structure of the object to be operated,a square box-like object size and orientation estimation method based on fusion information of multi-viewpoint point clouds is proposed.Firstly,using RGB-D data,SLAM algorithm is used to estimate and store some of the key poses in the robot's movement process.Then,the point cloud data from multiple perspectives are fused together according to the key pose estimation results to obtain a more complete object point cloud information.Then similar point cloud filtering and segmentation algorithms are used to extract the point cloud information of the target object.Finally,because the object point cloud information is abundant,the original RANSAC-based object edge straight line fitting method for the object contour is improved to recognize the orientation and size of the object.The experiment verifies the accuracy and robustness of the method.Aiming at the demand of transporting irregularly placed automobile objects in outdoor environments,a method for identifying the width and orientation of objects is proposed.The method adopts a three-dimensional laser sensor that is not interfered by outdoor sunlight,to construct a three-dimensional point cloud map of the entire environment using ICP based lidar SLAM.Then the target car point cloud is separated from the map,and the car point cloud is projected to the ground.Since the car point cloud is not a standard square,an optimizationbased method is used to estimate the outline of the car,and the width and orientation of the car is determined based on the estimated profile information.Experiments show that this method has good recognition effect on the width and orientation of the recognition car in the outdoor environment.