3D Scene Segmentation Based On Deep Point Cloud Convolution Neural Network DPCNet

In this paper,aiming at the problem of three-dimensional(3D)scene segmentation based on depth learning,the related technology and theoretical knowledge of 3D reconstruction algorithm are fully studied,furthermore the working principle and advantages of convolution neural network and point cloud depth learning network PointNet/PointNet++ are researched.At the same time,the shortcomings and defects of the point cloud data depth learning network directly acting on the data of point cloud for depth learning network are summarized.On this basis,an improved scheme for the fusion of two-dimensional image data features and 3D point cloud data features is proposed.After the improved model in this paper,the reconstructed scene is segmented by fine semantics,and the 3D object recognition is realized.The specific work of this paper is as follows:1.The RGB-D SLAM algorithm based on Kinect depth camera is used to reconstruct in the 3D scene.Firstly,according to the principle of depth camera,the color information and depth information of the scene are obtained,and then the feature points that are used to extract and register by the ORB algorithm,are extracted and matched for each frame of sample information captured.The PnP(Perspective-n-Point)method is used to estimate the pose of the camera,and the method of pose graph optimization was used for the back-end optimization.Finally,the BoW(Bag-of-Words)word bag model is employed for loop detection,and then gets the 3D scene reconstruction of point cloud data.2.On the basis of analyzing the depth learning and image segmentation technology,the semantic segmentation and object recognition in 3D scene are solved.In this paper,an improved network model that is deep point cloud convolution network(Deep pointcloud convolution network,DPCNet),is proposed on the basis of PointNet++ network model,and then depth learning algorithm based on 3D point cloud data is directly introduced.In the process of training with the proposed DPCNet,the image features extracted by two-dimensional CNN are fused with the 3D point cloud features,which enhances the matching between the adjacent data of the object point cloud.It improves the accuracy of the network model and the accuracy of semantic segmentation,and realizes the object recognition in the 3D scene.Based on the above work,the experimental environment of 3D scene reconstruction and the configuration of deep learning TensorFlow framework is established,meanwhile the semantic segmentation of 3D scene is carried out with PointNet++ and DPCNet respectively.The comparison results show that the improved network model has a high recognition rate in semantic segmentation.The test results of open data sets show that the semantic segmentation and object recognition under 3D scene reconstruction studied in this paper has certain theoretical value and good feasibility.