Deep Learning-based Pole Extractor for Long-term LiDAR Global Localization

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu | Master's thesis
Control, Robotics and Autonomous Systems
Degree programme
AEE - Master’s Programme in Automation and Electrical Engineering (TS2013)
Robust and precise localization is an essential requirement for an autonomous robot. Due to their distinctiveness and stability in the environment, pole-like objects such as trees, lamps, and traffic signs are frequently used as landmarks for long-term localization. This thesis presents novel deep learning-based methods that exploit pole information obtained by a LiDAR sensor to tackle both the local pose tracking and global localization problems. A fast pole extraction method based on geometric features has been developed for local pose tracking. All computations are executed directly on range images produced from LiDAR scans, which avoids explicitly handling the 3D point cloud. This range image-based method enables fast pole extraction from LiDAR data. The extracted poles are further used as pseudo labels to train a deep neural network for online range image-based pole segmentation to boost performance. The online pole extraction methods are integrated into the Monte Carlo localization (MCL) method as novel observation models to achieve accurate pose tracking. In order to achieve global localization, a robust LiDAR-based place recognition neural network is proposed. It retrieves several candidate places on the map given a query scan. Such candidates are then used as robot position hypotheses to initialize the MCL system. The proposed method achieves robust and accurate global localization by combining the pole-based pose tracking module with the LiDAR-based place recognition system. Thorough experiments on each module of the proposed method are conducted on different datasets with different types of LiDAR sensors, routes, and seasons. The evaluation results show that the proposed learning-based pole extraction method outperforms the geometric-based and baseline methods. After integrating the proposed LiDAR-based place recognition and pole-based pose tracking into the MCL, the proposed method achieves better localization performance than other state-of-the-art approaches and generalizes well across various datasets. Moreover, the proposed method runs online at the sensor framerate. In addition, the pole datasets made in this thesis and the implementation of the proposed LiDAR-based global localization approach are released to the public to facilitate further research.
Särkkä, Simo
Thesis advisor
Chen, Xieyuanli
localization, deep learning, autonomous driving, LiDAR, pole, range image
Other note