POINT CLOUD PROCESSING MODEL BASED ON DENSELY CONNECTED GEOMETRIC SHARED NEURAL NETWORK

When processing point cloud data, multi-layer convolutional neural networks have gradient loss, and it is difficult to obtain the global and local features of point cloud geometry at the same time. In order to solve the above problems and obtain sufficient contextual semantic information, a geometric shared neural network based on dense connections is proposed. In the multi-layer perceptron of the similarity connection module (Geometric Similarity Connection, GSC), each layer accepted the output from all previous layers as input, and its own feature map was used as the input of all subsequent layers. Its own feature map was used as input for all subsequent layers to effectively learn dense context representation and capture local and global geometric features. Experimental results show that the algorithm can effectively repeatedly aggregate similar and related geometric information and multi-scale semantics in point clouds, and can more effectively integrate the local structural characteristics of point clouds compared with the existing point cloud classification and segmentation algorithms, and further improve the accuracy of point cloud classification and segmentation.