Research On Object Recognition And Pose Estimation Based On Point Cloud

Facing to the demands of bin picking in modern warehousing logistics and industrial automation production,and in the case of 2D vision can not measure distance and estimate target pose correctly,studying object recognition and pose estimation algorithm based on point cloud.Using point cloud data to calculate precise poses of cylinders,boxes and complex-shape objects.In order to complete the task of grabbing cloth rolls,the cylinder segmentation and positioning method based on point cloud was studied.The problems of over-segmentation and over-growth in using the random sampling consensus algorithm were analyzed.The greedy clustering method was used after removing the boundary of segmented cylindrical point cloud to merge point clouds belonging to same cylinders.Using least squares method to calculate radius of cylinders,then edge points were projected to the centerline to calculate two endpoints of the centerline.Experiment results in laboratory and onsite show that proposed method can effectively segment the scene point clouds containing different size cloth rolls stacked randomly,and can get radius of cloth rolls and two endpoints of centerline.In order to pick up random stacking boxes,a plane-segmentation algorithm based on two classical segmentation methods was proposed,which solve the problem of cumulative normal deviation in grow process.Using greedy clustering method to fuse point clouds belonging to the same surface of the box,and rectangularity was used to screen surfaces unoccluded.Experiment results show that the proposed method can segment scene point clouds containing different size boxes stacked randomly.In order to recognize and locate objects have complex shape,a solution containing data preprocessing to hypothesis verification was proposed.Based on point pair feature,an improved method based was proposed to improve feature discrimination,dispersion and robustness to noise: In training phase,more discriminative point pairs were retained during down sampling,and point pair features were spread to 16 neighborhoods.In matching phase,using weight voting,and using local reference frames to verify candidate poses.The euclidean clustering method was used to screen poses belonging to different targets.After filtering model visible points,a coarse-to-fine ICP algorithm was used to calculate precise poses of the targets.Finally,unoccluded targets were screened by using overlapping rate.In Gaussian noise Experiments: when the standard deviation of Gaussian noise is 3% and 5% of the model diameter,the recognition rate is 98% and 78%,increasing 5% to 8%,compared to the original algorithm.Pose estimation tests and bin picking tests were carried out to complex scene point clouds,experiment results show that proposed method can calculate poses of multiple instances in 1.5 seconds,which can meet the demands of bin picking.