This study proposes a hybrid feature convolutional neural network (HFCNN) model for the complete description of three-dimensional (3D) point cloud features. The HFCNN confers sensitivity to the local, global, and single-point properties simultaneously by a feature vector space expansion. Wherein, a pointwise convolutional network sub-model realises the extraction of the local features by using a pointwise convolutional operator to process point cloud data directly. To consider the global properties of the point cloud, a central-point radiation model is constructed as an input of the feature layer in a nonnetwork form. Meanwhile, the single-point behaviour is characterised by the solo point coordinate information in the network. Within the constructed solo-local-global feature space, i.e. the fusion of single point feature, local feature and global feature, the HFCNN model can handle 3D point cloud data with unstructured and unordered properties. The HFCNN can be directly applied to the point cloud classification and segmentation without the modification of the CNN structure and training procedure. The experimental results have shown the effectiveness of the proposed model in prediction of class labels and point-by-point labels.

Voxel grid is widely used in point cloud segmentation due to its regularity. However, the memory consumption caused by high resolution restricts the performance of voxel grid. This paper proposes an improved voxel grid deep network (IVDN) model to represent more comprehensive point cloud features at the same resolution, thus improving the segmentation performance of point cloud. Firstly, the point cloud data are structured within a voxel bounding box to correspond with the three-dimensional(3D) convolution kernel, and a fixed number of point coordinates are selected to generate the point feature vector. Then, in order to consider the distribution characteristics, the reliability coefficient is used as an equivalent descriptor of the point cloud distribution density. Finally, a corresponding deep network is constructed to deal with the above features. Experimental results show that the proposed IVDN model can improve the mean classification accuracy and segmentation index mIoU(Mean Intersection over Union) effectively, with a 0.45% and 0.3% improvement on Shape net16 dataset respectively.

No abstract

No abstract