2D Map Estimation via Teacher-Forcing Unsupervised Learning

Existing global optimization mapping is prone to time-consuming parameter fine tuning of scan matching. In this work, we propose a novel map estimation method via cascading scan-to-map local matching with Deep Neural Network (DNN). Scan-to-map local matching firstly acts as a teacher to provide a coarse pose estimation, then a DNN is trained in an unsupervised learning fashion by exploiting the self-contradictory occupancy status of the point clouds. On the other hand, in order to cope with the mismatch problem caused by variable point number of a scan and fixed input size of DNN, a data hiding strategy is proposed. Experiments are conducted on three LiDAR datasets we collected from real-world scenarios. The visualization results of final maps demonstrate that our method, teacher-forcing unsupervised learning, is able to produce 2D occupancy map very close to the real world, which outperforms pure DeepMapping as well as ICP-warm-started DeepMapping. We further demonstrated that our results are comparable with those from traditional Bundle Adjustment (BA) method, without the need for parameter fine tuning.