Point clouds registration and semantic segmentation are two closely linked steps to build geometrically-accurate and semantically-rich 3D maps. They not only are the fundamental work in surveying and mapping tasks, but also contribute infrastructure to many downstream applications such as navigation and augmented reality. This paper addresses mapping and modeling indoor environments using mobile LiDAR systems (MLSs). We refer a scan or a laser scan to a single 360° sweep of a mobile laser scanner. Due to the inhomogeneous density of mobile laser scans, most existing simultaneous localization and mapping (SLAM) approaches suffer from unequally distributed feature points, which consequently leads to the odometry and mapping drift. To resolve this problem, we propose a novel mobile mapping framework, coined the hyteretic mapping. Essentially, we innovatively introduce a feature equalizer that seeks to balance the inhomogeneous feature points of buffered scans in a sliding window and make them uniformly distributed to the best possible extent. The feature points after equalization are more friendly to the feature-to-map registration process, so that the overall odometry and mapping drift can be alleviated. Moreover, we design a self-supervised motion prior generator network. This network consumes images of point clouds after spherical projection, and the single-track inference will directly provide an initial guess for the registration process. Finally, the point cloud map generated by the MLS is further segmented with information of interest. During semantic segmentation, we combine the existing method with two well-designed prior-information based modules to obtain more excellent performance.Extensive experiments were conducted on four aspects: mapping accuracy, trajectory correctness, study of different motion priors, and semantic modeling performance. In the MIMAP dataset, the proposed feature equalizer reduced the absolute mapping error by 35.6% from 5.51 cm (baseline) to 3.56 cm. In the NTU VIRAL dataset, we validated the motion prior generator and compared it with other three kinematic models. It helped our method achieve a 6.19 cm trajectory error which outperformed the state-of-the-art LiDAR-only and multi-sensorial fusion SLAM methods. The dataset for evaluating our semantic modeling method was acquired by the above mobile mapping method. It contained 100 training scenes with six common categories: floor, ceiling, wall, office door, escape sign, and window frame. Evaluation showed that the average mIoU of our method is about 7.1% higher than that of the baseline method. In addition to the public datasets of MIMAP and NTU VIRAL, our self-logged dataset, sample results and benchmark kits are also made publicly availble at https://github.com/chenpengxin/hysteretic_mapping.
Read full abstract