Hierarchical graph structure-based 3D point cloud map generation and entropy weighted localization method for navigation

Abstract

This thesis describes three basic frameworks for autonomous vehicles

Map generation, map management, and localization. First, this thesis describes a precise 3D map generation method based on a tilted 2D LiDAR scanner and demonstrates it. Conventional methods use expensive 3D sensors or can be applied only to a limited environment. However, this thesis shows the success of 3D precise map generation by using one tilted 2D LiDAR scanner installed in mobile robot platform, which guarantees lower expense compared to the conventional one. It is especially important to manage drift error of dead-reckoning while the 3D map is generated by using a push-broom-style LiDAR sensor. Firstly, therefore, this thesis proposes a 3D map building method that uses a robust interactive closest point matching algorithm and a drift error correction method based on pose-graph SLAM. A precise point-cloud map is generated with the hierarchical structure based optimization. The experiments were conducted in complex indoor and outdoor environments to verify the exact performance of proposed method. Second, this thesis pursue easy and effective grid map management method by using an image format. The generated point cloud from the map generation phase is transformed to three bytes image channels. Since all map information are stored into image forms, easier loading, saving, modifying, and management are possible. The proposed method is verified by converting city scale point cloud data. For the last, this thesis also proposes a robust vehicle localization method in urban area based on the prior point cloud. SLA (Singapore Land Authority) around one-north area in Singapore provides past high resolution point cloud collected six months ago. As the data are outdated, there exist many changed environmental aspects such as redrawn road markings, construction areas, and changing tree shapes. In response, this thesis proposes a novel fusion algorithm based on a particle filter using vertical and road intensity information for robust localization. While the state-of-the-art fusion algorithm focuses on optimization of the vehicle pose based on multiple measurements, this thesis gives another method: estimating a robust vehicle pose by considering the reliability of each feature from the prior map. Also, an efficient management strategy of a grid map including multi-layer vertical and road intensity information for real-time operation is proposed as well. A sensor system equipped on a vehicle consists of 32 channels of 3D LiDAR, IMU, wheel odometry, and they are used for the proposed algorithm. The ground truth was generated with the graph-structure optimization method by using RTK-GPS, wheel odometry, and ICP algorithm in post processing. The total data set for the demonstration is collected from 19.9km trajectory in urban area. The proposed approach has successfully performed an autonomous vehicle driving in urban area.