Research On 3D Object Recognition For Sparse Point Clouds

3D object recognition is a basic research in the field of computer vision.Its main task is to recognize objects from 3D data.It is widely used in many fields such as autonomous driving,virtual reality,and topographic mapping.Due to the inherent defects of point cloud acquisition equipment,the collected point cloud has sparseness.This sparse structure will lead to difficulty in capturing local relationships and poor representation ability,resulting in a significant drop in the recognition accuracy for deep networks.In order to solve the identification problem of the sparse point cloud,this thesis improves and optimizes the super-resolution method of sparse point cloud,feature extraction operator and recognition network structure.First,in view of the problem of the uneven density of point cloud generated by the super-resolution method,this thesis proposes a point cloud super-resolution method that combines attention mechanism and joint loss.Generate features with high correlation to realize super-resolution operation,and introduce homogenization loss on the basis of reconstruction loss to ensure that the generated point cloud is evenly distributed on the surface,and solve the clustering problem around sampling points in point cloud super-resolution.Secondly,in view of the difficulty of feature extraction due to the small amount of information in the sparse point cloud,this thesis designs a local relational convolution algorithm based on polynomial input and geometric prior.The key to this convolution is that it can learn from the relationships between points,and learn the geometric topological constraints between points,specifically,the low-dimensional geometry between the sampled point and all adjacent points in the local neighborhood of the reference point cloud.Therefore,a local geometric relationship description operator that does not depend on the order of points and rigid body transformation is constructed,to obtain more feature information,thereby improving the shape perception of point convolution and the robustness to the sparse point cloud.Finally,in view of the insufficient perception of local features in the single-scale recognition network model,according to the scale space theory,this paper designs a multi-scale structure.This structure is equivalent to connecting multiple hierarchical structures in parallel,realizing fusion and extraction of local features in multi-scale neighborhoods,and improving the perception of local features of the recognition network.By testing on standard point cloud data sets and comparing with other point cloud recognition algorithms,the experimental results show that the proposed method has higher recognition accuracy on sparse point cloud than other methods.It can achieve a better sparse point cloud recognition effect and provide a new solution to the sparse point cloud recognition problem.