Research On Point Cloud Registration Method Based On FMCW Scanning Radar

Point cloud data is composed of discretely positioned coordinates obtained by using optical sensors to scan the surfaces or structures of environments.These sensors are sus-ceptible to variations in lighting,environmental dust,and harsh weather conditions,all of which can degrade the quality of the acquired point cloud data.In contrast,frequency mod-ulated continuous wave scanning radar operates at lower frequencies with superior pene-tration capabilities,enabling the acquisition of stable quality point clouds even in adverse conditions.Consequently,applications based on radar point cloud data,such as localiza-tion and radar odometry,have been extensively researched.Radar point cloud registration is the core technology for processing radar point cloud data in these applications.Com-mon methods include direct radar point cloud registration and feature-based radar point cloud registration.However,these methods have limitations:(1)They do not adequately consider the effects of noise,multipath reflections,and radar beam spreading on the ac-curacy of radar point cloud registration;(2)The radar point cloud keypoint descriptors perform poorly in areas with complex pixel structures or insufficient texture information around keypoints,directly affecting the accuracy of feature matching and,consequently,the performance of radar point cloud registration;(3)They do not adequately address the sparsity of radar data,leading to low accuracy in feature matching and thus impacting the precision of radar point cloud registration.To address these limitations,this dissertation investigates high-precision point cloud registration methods based on FMCW scanning radar.The main research content and contributions include the following aspects:(1)Aiming at the problem of the noise,multipath reflection,and radar beam expan-sion,affecting the accuracy of radar point cloud registration,this research proposes a radar point cloud registration method based on local point cloud features.The proposed method begins with enhanced keypoint detection from radar scan data,improving the pre-cision of detected keypoints.The proposed method also constructs a low-dimensional radar point feature histograms descriptor combined with a coarse-to-fine point cloud reg-istration framework,improving the accuracy of radar point cloud registration.(2)Aiming at the problem that the steered binary robust independent elementary fea-tures descriptor constructed based on random sampling templates leads to a decrease in the discriminative power of the descriptor and the accuracy of feature matching when the pixel structure near the keypoint is complex,or the texture information is insufficient,this research proposes a radar point cloud registration method based on the rotation-aware brief(rBRIEF)descriptor.The proposed method firstly improves the keypoint detection method,on this basis,constructs the rBRIEF descriptor using sampling templates obtained based on the large number of training in the image dataset,effectively improving the de-scriptiveness and accuracy of the keypoint descriptor.Additionally,a coarse registration algorithm based on the maximum clique of random submatrices is proposed,which,com-bined with fine radar point cloud registration methods,significantly improves the accuracy of radar point cloud registration.(3)Aiming at the problem of characterizing and matching keypoints in the radar point cloud without considering the sparsity of radar data that leads to low feature match accuracy,this research proposes an image feature-assisted registration method for the radar point cloud.The method first converts the raw radar scan data into the correspond-ing image in the Cartesian coordinate system by interpolation,detects the features on the image,and performs feature matching by using the method in the image domain,which effectively improves the correct rate of feature matching.Additionally,a coarse registra-tion algorithm based on random sample consensus combined with geometric consistency constraints is proposed,which,combined with fine radar point cloud registration methods,enhances the accuracy of radar point cloud registration.The experimental results of point cloud registration on the FMCW scanning radar dataset indicate that the methods proposed in this thesis can mitigate the effects of noise,multipath reflections,and radar beam spreading on keypoint detection.These methods en-hance the descriptiveness of the descriptors and the accuracy of feature matching,thereby improving the registration accuracy of radar point clouds.