We propose a framework for odometry, mapping and ground segmentation using a backpack LiDAR system that achieves both real-time and low-drift performance. First, we present a spatio-temporal calibration method to carefully merge scans from the two laser scanners on a backpack. Second, we propose a feature extraction method which generalizes a point's geometrical characteristics as two groups (disjoint, continuous) and three types (edge, corner, plane). The extracted features are used in point cloud registration in the odometry and mapping tasks. Third, a fast ground segmentation method is customized for the backpack LiDAR system. Finally, we evaluate the proposed method in four datasets logged by the backpack across different scales and environments. Furthermore, NTU VIRAL dataset is used to benchmark our method quantitatively. Experiments show that our method consistently outperforms the state-of-the-art methods before using loop closure optimization and sensor fusion techniques.

Seamless localization service is regarded as an essential requirement towards smart city based applications. The performance of seamless localization is restricted by the limited coverage of indoor navigation database and disturbed results of the location sources in complex urban environments. This paper proposes a map-assisted seamless localization framework that uses the self-constructed navigation database and Bi-directional Long Short-Term Memory (Bi-LSTM) based signal quality control criteria (SL-SDBQ). Daily-life trajectories provided by the inertial odometry are optimized and calibrated by the map-originated indoor pedestrian network information for self-constructing the crowdsourced navigation database, and the Bi-LSTM is applied for quality evaluation of different location sources. In addition, the robust Kalman filter is adopted to integrate the information of inertial odometry, crowdsourced Wi-Fi fingerprinting, Global Navigation Satellite System (GNSS), and signal quality evaluation results, and the error ellipse is applied for map matching to further enhance the performance of multi-source fusion based seamless positioning. The experimental results prove that the proposed SL-SDBQ can achieve autonomous and precise indoor and outdoor positioning performance, and meter-level localization precision can be realized under the assistance of indoor map.

Loop closure detection is a critical yet still open technique to enhance the performance of Simultaneous Localization and Mapping (SLAM). In this paper, a very deep and lightweight neural network DeLightLCD is proposed to enable efficient loop closure detection. The architecture of the network contains two key modules: a siamese feature extraction module and a dual-attention-based feature difference module. A very deep but lightweight feature extraction network is designed to extract high-dimensional and discriminative features. The dual-attention-based feature difference network generates feature difference maps to identify the difference between a pair of LiDAR scans. Besides, a loop closure candidate rapid selection method is proposed to extract loop closure candidates from the LiDAR scan sequence. The workflow enables real-time loop closure detection in large-scale environments. This approach was evaluated on the open-source KITTI odometry benchmark and Ford campus dataset. The experimental results show that this method outperforms the state-of-the-art methods. Although the model was trained only on the KITTI dataset, it also demonstrated superior performance on the Ford campus dataset. In particular, the proposed approach is more lightweight and highly efficient than the current existing approaches.

A mobile mapping system (MMS) is a widely-used moving platform for the acquisition of rich geospatial information and requires sensor calibration to guarantee data quality. Global navigation satellite system (GNSS) sensor and light detection and ranging (LiDAR) are important parts of an MMS, and the extrinsic calibration between them can markedly improve its performance. A sole GNSS sensor is generally used in a portable MMS, which is the current trend. However, existing extrinsic calibration methods of GNSS/inertial navigation system (INS) are not applicable for the sole GNSS sensor because the former requires relative six degrees of freedom (DoF) pose output for both sensors and the latter can only provide 3-DoF positions. Here, we propose two sole GNSS sensor extrinsic calibration methods: one is a target-based direct calculation method and the other is a non-target-based automatic iterative method. The experiment results show the two methods have good repeatability, the error of the target-based method is below 3 cm, and the error of the non-target-based method is below 2 cm.

Maps are indispensable foundation for autonomous driving to provide prior knowledge of road structures. Constructing maps in a crowdsourcing manner has been a research hotspot in order to achieve lower map production cost and higher update frequency. Existing pipelines of crowdsourcing mapping construct scalar maps prior to on-cloud map vectorization. This leads to not only more demand of storage space and uploading bandwidth, but also the waste of computing resources because the on-cloud map vectorization is doing the similar work as the on-vehicle scene perception. To solve this issue, the letter proposes a novel, lightweight and efficient crowdsourcing mapping pipeline coined MapCVV. The pipeline directly consumes visual vectorized elements over the shelf provided by on-vehicle cognition and outputs vectorized map which can be used for localization, planning and navigation. Considering the lack of local rigidity in visual vectorized elements, traditional submap-level optimization will not be applicable. To cope with this challenge, we propose the novel element-level optimization that can achieve better global consistency among multiple runs when dealing with visual vectorized elements. Finally, we validate the practicability of the proposed MapCVV pipeline by conducting a 445 km real-world experiment. The experiment shows that our method outperforms others with regard to global consistency and absolute accuracy. The uploading bandwidth for crowdsourcing data is only 2.6 MB/km. The average processing speed is notably 1.84 s/km.