Research On Target Recognition Algorithm Of Ladar Range Image Based On Point Normal Pose Estimation

Laser radar(ladar)is a three-dimensional(3D)stereo sensor.Compared with traditional two-dimensional sensors,such as passive infrared sensor and visible light CCD,laser radar can simultaneously acquire intensity images reflecting target material properties and 3D geometric range images reflecting 3D space structure of targets.The two images are collectively called four-dimensional image(3D range image plus one-dimensional intensity image)and can provide rich target information.Ladar is a revolutionary technology for improving target recognition rate.In order to fully utilize the 3D target recognition capability of ladar,3D target point normals were explored with the point cloud model matching algorithm in this paper.A recognition algorithm was proposed to directly match range images with 3D models.With the acquired multi-view ladar range images,3D model reconstruction was firstly achieved and the model library was established.Then,through estimating pose angles of range images and converting the target poses,coarse matching was realized.Finally,fine matching was performed with iterative closest point(ICP)algorithm for the purpose of target recognition.The algorithm proposed in the study mainly includes three parts: 3D model reconstruction,point normal pose estimation,and ICP match.The point normal pose estimation is the core of the algorithm and the key technique for achieving direct 3D model matching.The 3D model reconstruction algorithm was firstly studied in two aspects: data registration and data integration.In data registration,the matching accuracy decline may be caused by manually setting screening threshold of pseudo-point pairs.In order to improve the matching accuracy,an improved ICP algorithm(IICP)was proposed to automatically select the screening thresholds based on the comparison results of various surface types in 3D range images.Compared with the existing improved ICP algorithms,the IICP algorithm improved the accuracy of registration data,as indicated by the experimental results.In data integration,the integration accuracy decline may be caused by manually selecting the detection threshold in point neighborhood discrimination(PND)algorithm.In order to improve the integration accuracy,an optimized PND algorithm(IPND)was proposed to automatically determine the detection threshold based on the distribution of local point density variance of 3D model after registration.The experiment results verified that the IPND algorithm could obtain the better integration results without overlap or hollow,compared with the PND algorithm.Secondly,a 3D model matching recognition method was proposed based on 3D point normal pose estimation(PEPPN)to recognize rigid non-articulated targets with arbitrary poses.Through estimating point normals of 3D target range image,filtering planar-point normals by point curvedness,classifying planar-point normals by adaptive K-means clustering algorithm,a series of programs of the vectors in each coordinate axis direction in the target coordinate system were determined with the minimal normals of intersection angles between normals as representative normals.Then the target 3D pose could be estimated and coarse matching was realized.Then ICP algorithm was used to complete the fine matching.The algorithm was verified in three aspects: range accuracy,signal-to-noise ratio(SNR),and occlusion rate.The verification results showed that the pose estimation accuracy and recognition rate could be improved by increasing the range accuracy and SNR while the occlusion rate was retrained to a certain degree.Finally,in order to recognize rigid articulated targets with arbitrary poses,a 3D model matching recognition method was proposed based on pose estimation of each part of the target.Through 3D pose estimation of main body with the PEPPN algorithm,segmenting the main body and the articulated part according to point histogram distribution,decomposition and pose estimation of the articulated sub-section according to 3D geometry characteristics(point histogram distribution,point distance metric,and linear characteristic),each part of the target was transformed into canonical pose and rough matching and fine matching were completed for target recognition.The method proposed in this paper achieved the direct matching between range images and 3D models through pose estimation.The method can solve the recognition problem of 3D objects with arbitrary poses,reduce the complexity of the algorithm,and reach a high recognition rate,indicating its good application prospect.