Point Cloud Semantic Segmentation And Modeling

In this paper,the point cloud data captured by Lidar is taken as the research object,and deep learning is used as a research tool to deeply explore the cognitive problems of driverless vehicles in driving scenarios.Effective knowledge of driving scenarios is an extremely important part of driverless driving.In the process of driving,driverless vehicles need to acquire effective information from the surrounding driving environment in real time,and then effectively understand the information,so as to pave the way for the subsequent path planning and navigation.Point cloud semantic segmentation plays an important role in the environment perception of driverless vehicles.It can provide a 3d semantic world that is easy to understand for driverless vehicles.Therefore,this paper uses the method of deep learning to build a basic segmentation network to segment the point cloud scene,and uses some measures to improve the segmentation effect of the network,and finally realizes the effective segmentation of the point cloud space scene.The main work of this paper is as follows:Firstly,this paper introduces the acquisition method of point cloud and the characteristics of point cloud.Then,based on the characteristics of point clouds and existing computing resources,a method for converting point clouds is proposed.Point cloud is the collection of a series of reflection point coordinates in space.It is a sparse and unstructured data structure.When three-dimensional convolutional neural network is used to process point cloud,it will lead to the strain of computing resources,which is not desirable for the existing computing platform.Therefore,this paper proposes to map the three-dimensional point cloud to the two-dimensional space,reduce the dimension of input data,and process the point cloud in the two-dimensional space,so as to reduce the burden of computing resources.Secondly,in order to effectively segment the mapped point cloud,this paper builds a basic segmentation framework on the basis of UNet,and makes some changes to the established basic network to enhance the segmentation effect of the network on objects in the point cloud scene.For example,Focal Loss is used to enhance the network's detection accuracy on small target objects,different attention modes are used to improve the model's utilization of information flow in the network,and in the initial stage of feature extraction,an information aggregation module is proposed to enhance the network's anti-interference capability against data noise.In addition,the Vortex Pooling module is used to ensure the resolution of feature map when increasing the network receptive field,and a novel semantic enhancement module is proposed to reduce the semantic gap between low and high level features in order to enhance the effect of feature fusion.Finally,based on the above improvements,this paper builds a complete framework for semantic segmentation of point cloud scenarios.The experimental results show that the proposed model can effectively segment the point cloud scene.