Generalized LOAM: LiDAR Odometry Estimation With Trainable Local Geometric Features

This letter presents a LiDAR odometry estimation framework called Generalized LOAM. Our proposed method is generalized in that it can seamlessly fuse various local geometric shapes around points to improve the position accuracy compared conventional and mapping (LOAM) method. To utilize continuous features for estimation, we incorporate tiny neural networks into iterative closest point (GICP) algorithm. These data association metric matching cost function using features. Experiments with KITTI benchmark demonstrate our reduces relative trajectory errors other methods.