Semantic Segmentation Method Of Road Point Cloud Based On Mobile Measurement System

The point cloud collected by the high-precision mobile Li DAR measurement system is a digital infrastructure for high-precision mapping,autonomous driving,digital twins,and other fields.Unlike indoor point cloud data,the road scene point cloud data collected by the mobile Li DAR measurement system at high speed is not dense and uniform.In the process of high-speed scanning,there is a large amount of missing and noisy road point cloud data.When collecting data at different speeds,the uniformity of point cloud data in different road sections also varies due to the different scanning density at different distances.At the same time,the mobile measurement system can quickly obtain the point cloud data of surrounding objects.However,the volume of the obtained point cloud data is enormous,and manual processing of the mobile measurement system’s point cloud data requires a long time.Therefore,the problem of automatically segmenting a large amount of uneven and sparse point cloud data in road scenes has become a research hotspot.With the advancement of technology and methods,convolutional neural networks(CNN)have achieved remarkable results in 2D object recognition,and more researchers have started to extend deep learning to the 3D field.Inspired by the Point Net++ network,we propose a new end-to-end deep learning network that utilizes road point cloud data collected by the mobile measurement system for semantic segmentation.This addresses the problem of low semantic segmentation accuracy in existing deep learning networks for the uneven and sparse point cloud data in mobile Li DAR system measurements(MLS).To solve the problem of low accuracy of deep learning networks in segmenting sparse and non-uniform road point clouds,we propose a module with an adaptive sampling radius.The Euclidean distance between the sampling center point and all selected points is calculated.Points with a distance larger than r2 are assigned as N.All points within the sampling radius are sorted in ascending order,and n points are selected as the sampling points.As there may be points with N values among the n points,i.e.,there are fewer than n points within the spherical sampling radius,we check whether the sampling points contain any points with N values.If so,we expand the spherical sampling radius and check again until we have enough points to learn high-dimensional features and enhance the features of the sampling points.To improve the efficiency of deep learning networks when performing semantic segmentation tasks and to reduce the training and testing time,we incorporate a channel attention mechanism for mobile Li DAR point cloud data into a deep learning network model.After obtaining the high-dimensional features of each point,we focus on the impact of information on the entire point cloud information in each feature channel.By obtaining the correlation between features and information on each channel,we implement a point cloud data channel attention mechanism to highlight the most critical information on each point feature channel.We first obtain a feature that reflects the overall point cloud,and then learn the attention weights of each channel on the point cloud features to obtain the point cloud features processed by the feature channel attention module.In order to train and test the deep learning network proposed in this paper,we build a deep learning road point cloud dataset based on a self-developed mobile Li DAR system.We select suitable road section point cloud data from different road scenes as the training dataset for the deep learning model.Based on the mobile Li DAR system we developed,the point cloud data of the mobile measurement system road dataset is based on the road data of Beijing’s surrounding roads,including the Daxing Airport Line,the Beijing-Taipei Expressway,the Beijing-Kaifeng Expressway,and the Beijing Ring Expressway.Finally,road point cloud semantic segmentation training and testing are conducted,and the overall accuracy of our method for mobile measurement system road point cloud segmentation is98.08%,with an overall m IOU of 0.73.The m IOU values for road,guardrails,signs,streetlights,and lane lines are 0.99,0.983,0.99,0.66,and 0.51,respectively.Experimental results show that our research can achieve accurate segmentation for uneven and sparse mobile Li DAR road point clouds.Compared with existing methods,our segmentation accuracy is significantly improved.