Semantic Segmentation Of Real Point Cloud Scenes Based On In-depth Learning

With the development of point cloud scanning backpack and point cloud data acquisition devices such as vehicle-mounted lidar sensors,point cloud data has been widely used in many scenarios,whether unmanned or industrial robots,virtual reality,geographic information systems and other fields,which require the technical support of point cloud semantics segmentation for large-scale real scenes.After a breakthrough in two-dimensional image segmentation,data-driven in-depth learning is also used for point cloud semantics segmentation.Compared with two-dimensional image semantics segmentation,three-dimensional point cloud data is discrete,unstructured and out of order,and has its own unique three-dimensional spatial location information.In order to effectively segmente point clouds,extracting spatial geometric structure features based on the spatial location information of point clouds,previous methods failed to make full use of the global geometric structure features of point clouds,or failed to take into account both global and local geometric structure features,resulting in inconsistent global structure or inaccurate local structure.In addition,point cloud data acquisition is more difficult than image data acquisition,and the lack of specific categories of point cloud sample data is also a problem in semantic segmentation.The improvement of the model network of deep learning by the previous methods makes the number of layers deeper,the amount of parameters more,and the amount of training data more required.In view of the above problems,the main research contents and innovations of this paper are as follows:1.Based on Rand LA-Net framework,a spatial structure feature,point box feature,is proposed for semantic segmentation.A network framework with codec-decode structure is designed,in which the global spatial and local neighborhood features of point clouds are learned using the geometric structure feature module during the downsampling process,and the full size feature map is restored step by step by resolution during the upsampling process for semantic segmentation,which integrates global and local features.This method effectively obtains the generalized and fine-grained geometric features of point clouds.Experiments on several datasets show that this method can effectively improve the accuracy of segmentation.2.A semantically segmented data augmentation algorithm for large-scale real point cloud scenarios is presented,including spatial search algorithm and spatial replacement algorithm:(1)Spatial search algorithm extracts local data scenes from original dataset scenes according to the principle of edge detection,performs various point cloud data augmentation processing,generates complete and non-morbid local data scenes,and compares the improvement of different point cloud data augmentation methods through experimental analysis.(2)The spatial replacement algorithm generates new datasets by replacing local data scenes from point cloud scenes.The network model can be used directly to train with the generated point cloud data or to solve the problem of uneven data distribution using samples with a scarce type of data.Experiments show that this method can generate "challenging" data,which enables the network model to effectively predict potential transformations of data and thus improve the accuracy of segmentation.