Research And Implementation Of 3D Point Cloud Target Segmentation Algorithm For Multi-grade Highway

With the development of LIDAR technology,three-dimensional point cloud data collection becomes easier,so three-dimensional point cloud data is increasingly used as environmental perception data in research and engineering fields.At the same time,the rapid development of deep learning technology makes three-dimensional point cloud target segmentation technology based on deep learning method has been widely used in automatic driving,assisted driving and other fields.Although,three-dimensional point cloud data acquisition and processing technology is very mature,lack of public road three-dimensional point cloud data sets has been the problem at home and abroad,which severely restricted the development of threedimensional point cloud target segmentation algorithm.At the same time,the completion of three-dimensional point cloud target segmentation task under large scenes such as highway environment is still faced with a large amount of data,complex environment and other problems,resulting in low efficiency and low accuracy of target segmentation.Therefore,starting from the practical application and existing difficulties of threedimensional point cloud processing,the following work is made in thesis:1.A complete collection and production scheme of three-dimensional point cloud data set of multi-grade highway was designed and implemented in the multi-grade highway environment around Chengdu,Sichuan Province,China.Through this data set collection and production scheme,we collected three-dimensional point cloud data of multi-grade highway environment about 500 km,and made three-dimensional point cloud public data set of multi-grade highway with effective about 180 km.The completion and disclosure of this data set lays a foundation for three-dimensional point cloud object segmentation algorithm in highway environment.2.A voxel-like preprocessing method and an improved Rand LA-Net model were proposed to solve the problems of low efficiency and low precision of multi-grade highway three-dimensional point cloud target segmentation algorithm.For high-grade highway,Map and K-D tree are used in thesis to realize voxel-like preprocessing method to reduce the amount of point cloud data and improve the efficiency of the algorithm.For low grade highway,the Rand LA-Net model is improved by using multilevel stacked residual blocks to improve the model accuracy.Experimental verification shows that compared with the baseline model Rand LA-Net,the target efficiency and accuracy are significantly improved in the self-made multi-grade highway three-dimensional point cloud data set by using the data preprocessing method and the improved model.3.The three-dimensional point cloud processing system of multi-grade highway is designed and implemented based on the three-dimensional point cloud object segmentation algorithm of multi-grade highway.The system realizes the basic function of three-dimensional point cloud pretreatment and three-dimensional point cloud identification of multi-grade highway.The system can be developed iteratively and widely used in three-dimensional point cloud processing field to solve the timeconsuming and laborious problem of manual three-dimensional point cloud processing.