Researches On Semantic Segmentation And Classification Based On Convolutional Neural Networks

With the emerging of low-cost sensors,acquiring 3D point clouds has become more convenient than before.Hence,extensive attention has been paid to the understanding of 3D point cloud in the computer vision field.Lots of applications in future life are highly closely related to it,such as autonomous driving,indoor navigation,human-computer interaction and virtual augmented reality.In 3D point cloud understanding,scene-level point cloud semantic segmentation and 3D object classification are two key problems.While these two problems can be solved by deep learning,difficulties and challenges remained due to order-less and nonstructural nature of point cloud.In this paper,based on convolutional neural networks,we propose a scene-level point cloud semantic segmentation algorithm and a 3D object classification algorithm to address the technical difficulties of scene point cloud semantic segmentation and 3D object classification tasks.For the scene-level point cloud semantic segmentation algorithm,since previous handcrafted features lack representational abilities,and traditional convolutional neural network architectures cannot be directly applied to the 3D point cloud,a local coordinate system plane convolution operation is proposed to directly process the point cloud.The proposed convolution method can effectively extract local structured information in the point cloud.The computation cost can be significantly reduced by using depth separable convolution.Using an encoderdecoder network structure akin to the U-Net structure of image semantic segmentation as the overall architecture of the point cloud semantic segmentation algorithm,with a jump connection between the encoder and the decoder,the model can use high-level semantic features and lowerlevel texture features of the network when making predictions.The network can achieve high prediction accuracy by learning from point cloud semantic segmentation dataset.For the 3D object classification task,a convolutional neural network model is designed by considering characteristics of point cloud,including order-less,invariance of rigid transformation,the structural correlation between points,and leveraging the planar convolution of local coordinate system.The proposed network randomly samples a set of points from the 3D object surfaces as input.It then uses the local coordinate system plane convolution to organize and associate the neighborhoods of the point,and extracts the local structural features of the point cloud.For learned local features of the point cloud,a global feature capable of generalizing the three-dimensional object is obtained by performing a maximum pooling operation on the features of all the points.For proposed network model,the adaptation ability of the network to the rigid transformation of the 3D object is enhanced by introducing a spatial transformation network module among the convolutional layers.The experimental results show that proposed network structure can significantly reduce model memory and overall computation while maintaining high accuracy,compared to previous voxel-based network model algorithm.