Training Samples Automatically Generation Of Laser Scanning Point Clouds Based On 2D Topographic Map And Deep Learning

As a real-time active device,laser scanning system can obtain large-scale and highprecision three-dimensional geographic information all day and all-weather,and has been widely used in power line patrol,disaster prevention and control,unmanned driving,smart city construction and other fields.Semantic annotation of point clouds is a basic task in the application of laser scanning data.In order to reduce the manpower and material resources consumed in the actual production,domestic and foreign scholars have carried out in-depth researches on the automatic classification method for laser point clouds.Deep learning algorithms based on their ability to extract high-level representations through compositions of low-level features,have achieved excellent performance in semantic labeling tasks of point clouds.A major factor that affects the result of deep learning method is the performance of training samples.Due to the problems of the large amount of data,complex topography and distribution of ground objects,the progressive processing point clouds by using deep learning needs massive training samples.In order to reduce the resources consumed in the generation of training samples,this paper proposes a solution based on the two-dimensional topographic map data and deep learning.Firstly,the laser point cloud classification based on twodimensional topographic map and deep learning are studied separately,and then the two research ideas are combined to solve the problem that is difficult to handle only using two-dimensional topographic map,and automatically generate training samples to meet the needs of deep learning.The main research contents are as follows:(1)Two-dimensional(2D)topographic map contains rich geographic location and semantic category information,which will greatly improve the accuracy and efficiency of laser-point cloud data classification.To solve this problem,this paper proposes a classification method for laser scanning point clouds based on 2D topographic map.Firstly,unsupervised filtering is used to roughly classify point cloud data into three categories: ground points,just above ground points and above ground points.According to these initial height labels,the corresponding polygon features in the 2D topographic map are selected for data registration,and the category labels were assigned.This will greatly reduce the misclassification caused by the registration of 3D data and 2D data.For the point cloud data which cannot be processed by polygon features,an unsupervised segmentation method is used to transform it into semantically consistent segments.For these segments,point features in the 2D topographic map are used for registration in an iterative way,and category labels are assigned.Finally,the rule-based method is used to process the remaining unclassified point clouds.The proposed algorithm has been tested on the datasets in Rotterdam city,containing arounds 30 million points.The overall accuracy is about 90% in the accuracy evaluation system based on the 2D topographic map,without feature extraction and training steps.(2)Classification of point clouds is the basic task in laser scanning point clouds processing.It is quite a challenge when facing the complex observed outdoor scenes and the irregular point distributions.In order to reduce the computational burden of point-based classification method and improve the classification accuracy,we present a segmentation and multi-scale convolutional neural network based classification method.Firstly,a feature image generation method is used to transform the 3D neighborhood features of a point into a 2D image.Then,a three-step region growing segmentation method is proposed to reduce both under-segmentation and oversegmentation.Finally,feature images are treated as the input of a multi-scale convolutional neural network for training and testing tasks.In order to get performance comparisons with existing approaches,we evaluate our framework on International Society for Photogrammetry and Remote Sensing Working Groups II/4(ISPRS WG II/4)3D labeling benchmark tests.The experiment result which achieves 84.9% on overall accuracy and 69.2% on average F1 scores has a satisfactory performance over all participants.(3)Training samples play an important role in the supervised classification.Most of the training samples are generated by manual labeling,which is time-consuming.Combining the ideas of the first two studies,to reduce the cost of manual annotating for laser scanning point clouds,we propose a framework that automatically generates training samples using a 2D topographic map and deep learning method.In this approach,input point clouds,at first,are separated into the ground part and the nonground part by a DEM filter.Then,a point-in-polygon operation using polygon maps derived from a 2D topographic map is used to generate initial training samples.The unsupervised segmentation method is applied to reduce the noise and improve the accuracy of the point-in-polygon training samples.Finally,the graph convolutional neural network is used to further purify training samples,so as to solve the problem of point clouds classification which is difficult to deal with only using 2D topographic map.A comparison with the point-based deep neural network Pointnet++(average F1 score 59.4%)shows that the segmentation based strategy improves the performance of our initial training samples(average F1 score 65.6%).After adding the intensity value in unsupervised segmentation,our automatically generated training samples have competitive results with an average F1 score of 74.8% for ALS data classification while using the ground truth training samples the average F1 score is 75.1%.The result shows that our framework is feasible to automatically generate and improve the training samples with low time and labour costs.