Design And Implementation Of 6D Pose Estimation System Based On Point Cloud Registration

The 6D pose estimation technology has been widely applied in practical fields such as unmanned driving,robot perception,virtual reality,and industrial assembly.With the rapid development of high-precision sensors such as LiDAR and Kinect,three-dimensional point clouds have become an important form of characterizing the three-dimensional world.By registering 3D point clouds,the pose transformation matrix between objects represented by point clouds can be calculated,becoming an important means of 6D pose estimation.However,in low overlap 3D point cloud point to point scenarios,how to obtain robust point to point correspondence,improve the accuracy of point cloud registration,and improve the accuracy of 6D pose estimation has become a key issue.This article conducts the following research on this issue:(1)This paper proposes a feature fusion method based on channel selfattention mechanism to address the issue of feature extraction modules in PointDSC network that ignores the correlation between features of different dimensions.Firstly,a non-local feature extraction module guided by channel self-attention mechanism was designed.Then,by fusing the matching pair features extracted from this module and the SCNonlocal module,better geometric features were provided for each input matching pair.The experimental results show that the improved algorithm has improved the registration recall(RR),interior prediction accuracy(IP),interior recall(IR),and comprehensive evaluation index(F1 measure,F1)by 1.54%,0.82%,0.35%,and 0.62%,respectively,compared to the baseline model,verifying the effectiveness of the improved method.(2)In response to the problem of ineffective elimination of incorrect matching pairs during the sampling stage of the matching pair in PointDSC network,this paper proposes a quadratic filtering strategy based on spatial angle invariance,which utilizes the invariance of spatial angle after rigid body transformation to achieve quadratic sampling of the seed point matching pair in the subset.Compared with the baseline model,this method improved registration recall(RR),interior point prediction accuracy(IP),interior point recall(IR),and comprehensive evaluation index(F1)by 4.38%,2.34%,4.17%,and 3.23%,respectively,verifying that the secondary screening strategy can more effectively eliminate incorrect matching.By combining the improved Nonlocal feature fusion module based on channel self-attention mechanism and the secondary filtering strategy based on spatial angle invariance,the network improved by 5.61%,2.89%,5.52%,and 4.18%on the above four indicators,respectively.Compared with mainstream algorithms of the same type,it has higher registration accuracy and better generalization ability in low overlap scenarios.(3)Due to the fact that the coarse registration algorithm can only complete preliminary pose estimation,this paper further obtains a more accurate pose transformation matrix through the G-ICP fine registration algorithm.In response to the problem of high time consumption in the GICP algorithm,this article applies the KD tree nearest neighbor search algorithm based on BBF to the G-ICP algorithm.The experimental results show that the method can effectively shorten the time consumption of the G-ICP algorithm.Finally,based on the above research,this paper designs and implements a 6D pose estimation system based on point cloud registration algorithm.The system provides users with functions such as point cloud preprocessing,key point extraction,pose estimation,and visualization to meet practical needs.