Research On Vehicle-borne LiDAR Pole-like Objects Point Cloud Classification Based On Deep Learning

Pole-like objects(street trees,street lights,road traffic signs,line poles,etc.)are an important part of road scenes.Their automatic and intelligent data collection,information extraction and classification are of great significance for building smart cities and smart transportation.As a cutting-edge spatial information acquisition technology,vehicle-borne LiDAR technology can quickly obtain the three-dimensional spatial information of urban features during vehicle driving,and the dense point cloud accurately describes the spatial form of various features,providing effective technical support for pole-like objects data acquisition.However,due to the limitation of data acquisition mode,vehicle-borne LiDAR records the point cloud information of buildings,cars,pedestrians,roads,rod-shaped features and so on.How to extract and classify pole-like objects efficiently and accurately has always been a research hotspot of scholars at home and abroad.With the in-depth study,convolutional neural network is outstanding in point cloud classification.Based on the road scene point cloud data obtained by vehicle-borne LiDAR,this paper focuses on point cloud filtering,pole-like objects extraction,pole-like objects classification based on deep learning network model.The main research results are as follows:1.Ground point cloud filtering based on improved RANSAC algorithm.KD tree is used to divide the point cloud space and calculate the point cloud normal vector and the ground point cloud is extracted.Traverse the roughly extracted ground point cloud to obtain the average elevation,and use it to obtain the ground point set.Randomly select a point from the ground point set as the seed point,use the FPS algorithm to select the remaining two seed points and fit the plane model,bring the point set data into the plane model,judge whether the distance from each point cloud data to the plane model meets the preset threshold,count the ground points,and repeat this step until an optimal plane model is fitted,Finally,the optimal plane model is used to filter the ground points accurately.The total error of the improved algorithm is 2.99%,which is 2.55% lower than the original algorithm.2.Construction of pole-like objects dataset based on binary image method.The PCA algorithm is used to project the point cloud onto the plane and generate the original image according to the boundary points;Then the mapping relationship between the point cloud and the original image is constructed,and the point cloud is transformed into a binary image;Then,the binary image is expanded,the connected domain is marked and the boundary contour is extracted;The point cloud segmentation is realized by the position relationship between the point cloud and the boundary contour line;Generate the minimum bounding box for the segmented features,calculate the elevation,projected area and other information of the point cloud in the bounding box,and set a reasonable threshold according to the geometric characteristics of different features,so as to realize the effective extraction of pole-like objects point cloud.Finally,pole-like objects were normalized and sampled down to make the data set required by the experiment.3.Pole-like objects classification is realized based on improved Point Net.The Point Net classification network cannot obtain the local features of the point cloud.To solve this problem,a point cloud local feature extraction layer is added on the basis of the original network to improve the classification accuracy of the network by obtaining the local features of the point cloud.After the original point cloud data is processed by T-net network,the corrected data is sent to the local feature extraction layer.FPS algorithm is used to select a group of points from the point cloud data as the centroid of the local point cloud,KNN(k-nearest neighbors)algorithm is used to obtain the adjacent point clouds around the centroid points to construct the local area,the local point cloud is sent to the multi-layer perceptron to learn the local features,then the maximum pool operation is used to obtain the local features,and finally the local features are spliced with the global features,the classification results are processed and output by multi-layer perceptron.The classification accuracy of the improved network is 98.4%,which is 1.8% higher than that of the point net classification network.