AGCV-LOAM: Air-Ground Cross-View based LiDAR Odometry and Mapping

We propose an air-ground cross-view based LiDAR odometry and mapping method, AGCV-LOAM, which uses satellite images as prior information to mitigate the accumulated error. The system consists of a LiDAR SLAM method and an air-ground cross-view pose correction neural network, which is used to estimate the accumulated error. The neural network takes as input a LiDAR gird-map and a satellite image patch, and output the pose correction value which is added to the factor graph to perform pose optimization. We evaluate our method against baseline methods using the KITTI dataset and experimental result shows that our method is able to mitigate the position error of the original SLAM method. Besides, our method also outperforms other baseline matching method.