3D Map Optimization with Fully Convolutional Neural Network and Dynamic Local NDT

Due to multi-path effects, GNSS-based localization methods are not always reliable in urban transportation scenes. To solve this problem, matching-based methods, which compare the real-time sensor data with a prior map, are widely used for urban autonomous driving. In these methods, high-precision noise-free 3D map plays an essential role in vehicle localization. This paper proposes a 3D map optimization framework to generate such map with high efficiency and low memory consumption. First, a deep learning based method is designed to automatically filter out non-map objects in mobile laser scans during mapping. Then, a method named dynamic local NDT is introduced for mapping and localization to improve efficiency and reduce memory usage. Furthermore, a road segmentation method is exploited for further optimization. The proposed framework only relies on LIDAR and GNSS-INS, which makes it simple and easily conducted. The mapping and positioning experimental results show that the proposed framework outperforms the conventional NDT method.