Research On Object Recognition In 3D Point Cloud Scene Based On Feature Description

Target recognition has always been the core topic in the field of computer vision and pattern recognition.It is widely used in robot navigation,remote sensing and face recognition.In recent years,2D image target recognition technology continues to develop and has become a relatively mature field.Compared with 2D images,3D point cloud provides rich spatial information and is less disturbed by ambient light,so it can judge the spatial pose of the target more accurately.On the other hand,cheap 3D acquisition devices are developing rapidly and maturing continuously,making it easier to obtain 3D point cloud data.Based on the above advantages,target recognition based on 3D point cloud has gradually become a research hotspot in academia and industry.3D target recognition algorithms are mainly divided into two categories:recognition methods based on global feature description and recognition methods based on local feature description.The recognition method based on global feature description describes the global features of the scene,ignores the target surface shape and spatial information,and is difficult to identify the target and its pose in complex scenes such as occlusion and stacking.The recognition method based on local feature description is to extract local features in the specific neighborhood of key points,which is more suitable for the recognition of partial visible and incomplete object surface point clouds.However,it is not robust to the interference problems of noise and low resolution of target surface point clouds in actual complex scenes.For the problem of 3D point cloud target recognition in the above complex scene,this topic collects the original point cloud information through the depth camera,and designs a solution of point cloud segmentation,point cloud local feature extraction and pose registration.The main work of this subject can be divided into the following three points:(1)Aiming at the stability and repeatability of local feature description,this paper proposes two feature description methods based on local reference frame(LRF),both of which are designed based on the idea of spatial voxel segmentation.Firstly,the zaxis of LRF is calculated by weighted covariance matrix analysis,and then the x-axis is determined by the addition of the plane projection of the neighborhood point weighting vector,and the Y-axis of LRF is obtained by the cross multiplication of xaxis and z-axis.The experimental results show that the improved LRF method has excellent repeatability and stability.For the descriptor method,the first one is to segment the neighborhood of the key point ball based on the improved LRF,and output the label value of the voxel as the feature information;The second is a description method based on voxel label homogenization,which takes the external cube of the neighborhood of the key point ball as the spatial voxel,evenly segments it,and considers the eigenvalues of each voxel and its adjacent voxels,which enhances the stability of the description.The results show that the two feature description methods are better than the current advanced description methods.The results of registration experiments further verify the performance of the description methods.(2)For complex scenes with occlusion and stacking,this paper presents a combined point cloud segmentation algorithm.The algorithm is mainly composed of super voxel over segmentation method and region growth method based on concavity and convexity.Firstly,the scene point cloud is divided into super voxel blocks through super voxel over segmentation,and then according to the concave convex relationship between each super voxel block,the region growth clustering is carried out with the seed of the super voxel block as the center until complete segmentation is realized.Through this segmentation algorithm,the point cloud blocks in the scene can be divided into specific target point clouds,which reduces the interference of occlusion and stacking on target recognition to a certain extent.(3)For the target registration and pose calculation in the scene,the template matching and registration methods are designed according to the feature description method proposed in this topic.The features of the template are extracted by the feature description method based on voxel homogenization,and the nearest neighbor matching is carried out with the features of the target point cloud to classify and roughly register the target point cloud;Then,according to the characteristics of key points,the matching point pair is found to calculate the coarse registration attitude;Finally,the iterative nearest point algorithm is used to achieve accurate registration and obtain accurate rigid transformation matrix.Through the registration process,the accurate pose of the target in the world coordinate system is accurately calculated.