3D LiDAR Map Compression for Efficient Localization on Resource Constrained Vehicles

Large scale 3D maps constructed via LiDAR sensor are widely used on intelligent vehicles for localization in outdoor scenes. However, loading, communication and processing of the original dense time consuming onboard computing platform, which calls a more concise representation to reduce complexity but keep performance localization. In this paper, we propose teacher-student learning paradigm compress point cloud map. Specifically, first find subset points with high number observations preserve performance, is regarded as teacher map compression. An efficient optimization strategy proposed deal massive data With supervision compressed map, student model built by training random forest fed geometric feature descriptors each point. As result, able without referring expensive numerical optimization. Additionally, incorporating features, innovative can be generalized other new while no re-training required. We conduct thorough experiments multi-session dataset KITTI demonstrate effectiveness efficiency paradigm, comparison compression methods. The final results show that learned achieve comparable based at same time.