Object Classification Based on 3D Point Clouds Covariance Descriptor

摘要

We introduce a new covariance descriptor combining object visual (color, gradient, depth, etc.) and geometric information (3D coordinates, normal vectors, Gaussian curvature, etc.) for mobile robot with RGB-D camera to deal with point cloud data. The improved mismatching correction algorithm is applied in the feature point mismatching correction of 3D point cloud, and then this correction algorithm combined with the classification framework for the dictionary learning is applied in the object recognition of 3D point clouds. This descriptor is able to quickly match the feature points of the point clouds in the surrounding environment and realize the function of object classification. Experimental results show that this descriptor has an advantage of the compactness and flexibility compared with the previous descriptor, and greatly reduces the storage space.