Research On 3D Point Cloud Upsampling Technology Based On Deep Learning

With the increasing demand for 3D scene application,3D point cloud,a 3D data expression format,has attracted increasingly attention.However,the point cloud obtained by Lidar is often sparse,non-uniform and noisy,which will greatly limit the use of point cloud data.Therefore,how to obtain uniform,dense and clean point cloud by upsampling point cloud is very important.In recent years,with the rapid development of deep learning in the field of computer vision,scholars began to deal with the point cloud problem by using neural network.Combined with the above background and requirements,this thesis proposes a new method of 3D point cloud upsampling based on deep learning,and designs a lightweight and efficient point cloud upsampling network model to better solve the problem of 3D point cloud upsampling.The main contributions of this thesis are as follows:A new method based on neural network is proposed to realize point cloud upsampling.The network model is composed of three modules: point cloud feature extraction,perturbation learning and coordinate reconstruction.Through end-to-end learning,the network model can effectively upsample sparse,non-uniform and noisy point cloud,obtaining dense,uniform and clean point cloud.Inspired by the idea of surfaces differential geometry,an adaptive 2D perturbation learning method is proposed.The 2D perturbation value of the upsampled point relative to the input point is adaptively learned through the neural network,and then the 3D coordinate residual value is obtained,which is used to reconstruct the final upsampled point cloud.In order to better train the model and improve the upsampling effect,we have made a richer data set based on existing work,including 117 3D models and 36000 training data pairs.The experimental results show that the data set covers more features and contents,which is more conducive to the training of upsampling network.We have made intensive experiments and verification of the proposed method.The results show that compared with the existing work,quantitative evaluation metrics such as Chamfer Distance and P2 F Distance of the proposed method exceed the current work,and the qualitative visualization effect is also significantly better than the current work.In addition,proposed network model is only 5.01 MB,which is much smaller than the network model of existing methods,and has more practical potential.