Robust Feature-based LIDAR Localization and Mapping in Unstructured Environments

Abstract

In robotics, simultaneous localization and mapping (SLAM) is a fundamental capability for autonomous mobile robots. This thesis deal with the problem of mobile robot localization and mapping in human-made environments. The contribution of this work is an innovative SLAM method based on robot odometry and LIDAR features (both keypoint and descriptor). The presented method does not requires an initial configuration of the environment and therefore can be adopted wherever the robot can operate. Feature-based approaches are a class of methods well studied in computer vision and 3D point clouds processing, but relatively new in 2d range sensing. The proposed LIDAR keypoint detector, named FALKO, with two novel descriptors, BSC and CGH, have been designed to provide stability and repeatability in feature-based laser scan matching. FALKO achieves higher repeatably score and extracts less ephemeral points than the other state-of-the-art keypoint detectors. Moreover, the precision-recall curves of the proposed descriptors are consistent with the achievable results obtained from computer vision and laser scan data descriptors. This thesis also illustrated novel loop closure methods based on FALKO keypoints and a novel feature signature, named GLAROT, and compared their performance in both offline and online localization and mapping problems with state-of-the-art signature algorithms. Results show that the FALKO detector combined with GLAROT signature and point-to-point association outperforms the previously proposed approaches. In this thesis, a novel automatic calibration method that simultaneously computes the intrinsic and extrinsic parameters of a mobile robot compliant to the tricycle wheeled robot model, which is a common kinematic configuration of industrial AGVs, has been proposed. The calibration is performed by computing the parameters better fitting the input commands and the sensor egomotion estimation obtained from the sensor measurements. Finally, an application of LIDAR featured-based localization for industrial AGVs, along with the automatic calibration method evaluation, have been illustrated. Results show that the feature-based approach achieves performance comparable to artificial landmark localization and compliant with requirements of reliable navigation. Both static and dynamic feature mapping have been tested showing that a static map of FALKO features performs like the reflectors counterparts.