| Abstract |
With the development of LiDAR detection technology, LiDAR point clouds have been applied in many fields such as surveying and mapping production, autonomous driving, and smart cities. However, fundamental researches on point cloud recognition, understanding, and analysis have encountered bottlenecks, seriously restricting the development of point cloud applications. The point cloud classification is an effective point cloud recognition technology that can achieve object classification, object detection, and semantic segmentation of point clouds. The point cloud classification method based on the feature learning network in this study consists of the feature encoding method, feature learning network, and backbone network. The feature encoding method can convert sparse and variable density point clouds into dense 3D voxels through voxel partitioning, point cloud grouping, random sampling, and KNN encoding; The feature learning network can transform unordered point clouds into regular and ordered grid level features through voxel feature encoding layer and bird's-eye view projection; The backbone network is composed of stacked hourglass networks, which can efficiently learn grid level features. In this paper, 2019 Data Fusion Contest dataset from IEEE Geoscience and Remote Sensing Society is used for experiments. The results of these experiments indicate the method of this paper can effectively classify point clouds and has a higher classification accuracy rate than other popular methods. © 2023 IEEE. |
