Research On Data Conversion And Filtering Methods Of LiDAR Point Cloud Based Convolutional Neural Network

As an activate remote sensing measurement method for acquiring a wide range of geographic information,Light Laser Detection and Ranging(LiDAR)technology is not affected by factors such as terrain,weather,etc,and gradually becomes an important method to obtain geographic information data.As the product of LiDAR technique,point cloud has high precision,high density and huge data.Point cloud data is the product of LiDAR technology,which has high precision,high density and huge data,and high quality point cloud data is the key guarantee for the production quality of subsequent related map data products.At the same time,with the continuous maturity and development of deep learning technology,research combining deep learning with other fields is also a hot issue at present.Therefore,for the filtering processing of LiDAR point cloud data,this paper focuses on the subject of using deep learning technology to perform LiDAR point cloud filtering processing,and has carried out the following work.First,in this paper,four classic airborne LiDAR point cloud filtering algorithms are described,and the concepts and theories of convolutional neural network and deep learning are introduced.Next,two common conversion algorithms of point cloud in deep learning are introduced: ” point cloud to image” and ” point cloud to voxel”,based on these two methods,this paper made some improvements and a 3D LiDAR point cloud conversion algorithm are came up with.Based on the same experimental environment,three algorithms are compared.Experiments show that the 3D-LiDAR point cloud conversion algorithm is obviously superior to the point cloud to voxel method in processing speed.Compared with the method of point cloud to image,the processing speed is slower,but the difference of required time is not significant.Finally,the proposed improved algorithm is used to perform point cloud conversion.Experiments on airborne LiDAR point cloud filtering methods for deep learning are performed using conventional AlexNet and improved AlexNet,respectively.The experimental results show that the conventional AlexNet-based airborne LiDAR point cloud filtering method achieves 92.97%,and the improved AlexNet-based airborne LiDAR point cloud filtering method achieves 96.86%,indicating that the improved AlexNet model has obvious advantages.At the same time,a contrastive analysis is carried on based on the same experimental data set.The results show that the error rate of class I error based on improved AlexNet airborne LiDAR point cloud filter method is 1.22%,the error rate of type II error is 1.92%,and the error rate of type II error is significantly lower than other methods(2.26%,7.46%).Thus,the method proposed in this paper has certain advantages of precision in the recognition and extraction of non-ground points,and it is feasible to carry out airborne LiDAR point cloud filtering using deep learning technology.