We present a new fast Ground Decoupled 3D Lidar Odometry (GND-LO) method. The particularity of GND-LO is that it takes advantage of the distinct spatial layout found in urban settings to efficiently recover the lidar movement in a decoupled manner. For that, the input scans are reduced to a set of planar patches extracted from the flat surfaces of the scene, found aplenty in these scenarios. These patches can be labeled as either belonging to the ground or walls, decoupling the estimation into two steps. First, the ground planes from each scan, clustered from the ground patches, are registered. Then, the motion estimation is completed by minimizing the distance between the wall patches and their corresponding points from the other scan, whose pairing is iteratively updated.
GND-LO has demonstrated to perform both precisely and efficiently beating state-of-the-art approaches. Concretely, exper- iments on the popular KITTI dataset show that our proposal outperforms its competitors by reducing the average drift by 19% in translation and 4% in rotation. This is achieved by running in real-time without needing GPU or optimized multi-threading, as it is commonplace in the literature.
