Fast and lightweight loop closure detection in LiDAR-based simultaneous localization and mapping

Abstract

Over the last decade, complex urban environments raised higher demands on geographic information data collection. Traditional data collection methods gradually fail to meet the growing efficiency, completeness, flexibility, and safety requirements. The advent of mobile mapping systems (MMS) filled these gaps but it has also brought new challenges to data processing. The processing of mobile measurement data requires automated, accurate, and efficient algorithms, which have been the hottest research topics in many relevant fields. In the thesis, loop closure detection (LCD), as one of the core problems of Simultaneous localization and mapping (SLAM) will be studied in depth. Focusing on LCD in indoor environments, a fast and compact algorithm is proposed utilizing comprehensive descriptors extraction and machine learning. Besides, a novel double-deck loop candidate verification strategy is proposed to validate loop candidates and reject false positives. As for outdoor large-scale environments, point clouds do not exhibit significant structural and regular geometric characteristics. Thus, the deep learning model is utilized to mine advanced and high-dimensional features. A very deep and lightweight neural network DeLightLCD is proposed to enable efficient LCD. The framework contains two key modules: a feature extraction module and a feature difference module. Depth-wise separable convolution (DSC) and batch normalization (BN) are utilized to ensure that the network is lightweight and trainable. In practical use, the generalization and flexibility performance of LCD algorithms are affected by sensor changes and indoor-outdoor environmental changes. Thus, DeLightLCD++ is proposed to address these problems. The improvement of DeLightLCD++ is threefold. (1) A novel data presentation method encoding measurement distance and azimuth angle information is used to reduce the effects of sensor changes. (2) The architecture of the network is also adjusted to ensure that the algorithm is rotation invariant. (3) A loop candidate fast search method is used to suppress the computation cost and time cost increase due to ultra-long measurement distance. After loop closures are detected, the results will be utilized for the pose optimization to eliminate accumulative errors in LiDAR odometry (LO). An enhanced graph optimization strategy based on LCD results is utilized in this thesis. Besides, three types of loops in graphs, detected loop closures, pseudo loop closures, and enhanced loop closures are introduced. Then, experiments are conducted to study factors affecting trajectory optimization performance. Finally, some guidance is given on fieldwork and data processing of the mobile mapping backpack system. The proposed methods are evaluated on open-source datasets and in-house datasets. The in-house datasets are captured by a self-designed mobile mapping backpack system. The backpack is equipped with two multi-line laser scanners. Results show that the LCD algorithms are superior to state-of-the-art algorithms in precision, time efficiency, generalization performance, and flexibility. The optimization method could effectively improve the LiDAR odometry results and enable a consistent map result. In sum, this thesis focuses on LCD and optimization for LiDAR-SLAM. The three LCD algorithms presented in the thesis aim to solve LCD problems in indoor and outdoor large-scale scenes. The experiments exhibit the effectiveness and superior performance of the proposed algorithms. The work presented can be implemented in LiDAR-SLAM for surveying and mapping. Furthermore, it could be used for autonomous driving, high-definition maps, and urban 3D modeling.