LiDAR-based self-localization using an adaptive machine learning method

In this paper, we present our approach for a convolutional neural network (CNN)-driven feature extraction from 3D point cloud data that can be used for self-localization with respect to an existing feature map of the terrain. We use an image-based CNN method from the literature to extract local features from transformed point cloud data. We introduce a method for position estimation with the retrieval results of one 360° point cloud and compare the results of this LiDAR-based technique with those from the image-based model. Furthermore, we investigate how the use of occupancy grids generated from point clouds can be used to refine the feature map and thus achieve similar localization results with a lower reference data density for map generation.