Page 24 - 卫星导航2021年第1-2合期

El‑Sheimy and Li Satell Navig (2021) 2:7 Page 14 of 23

generate a database. Finally, the real-time measured fea- Learning (DRL) (Li et al. 2019c), and Gaussian Process
tures are compared with those in the database to localize (GP) (Hähnel and Fox 2006), have also been applied.
the device. With the rapid development of machine-learning tech-
According to the dimensions of measurements and the niques and the diversity in modern PLAN applications,
database, database-matching algorithms can be divided database matching has been attracted even more atten-
into the 1D (measurement)-to-2D (database) matching, tion than geometrical methods. Te database matching
the 2D-to-2D matching, the 2D-to-3D matching, and the methods are suitable for scenarios that are difcult to
3D-to-3D matching. In the 1D-to-2D matching, the real- model or parameterize. On the other hand, the inconsist-
time feature measurement can be expressed as a vector, ency between real-time measurement and the database is
while the database is a matrix. Such a matching approach the main error source in database matching. Such incon-
has been used to match features such as wireless RSS sistency may be caused by the existence of new environ-
(Li et al. 2017) and magnetic intensity (Li et al. 2018). ments and varying environments and other factors. Te
Examples of the 2D-to-2D matching are the matching survey paper (Li et al. 2020a) has a detailed description of
of real-time image features (e.g., road markers) and an the error sources for database matching.
image feature database (e.g., a road marker map) (Gruyer
et al. 2016), and the matching of 2D LiDAR points and Multi-sensor fusion
a grid map (de Paula Veronese et al. 2016). By contrast, Te diversity and redundancy of sensors are essential to
the 2D-to-3D matching is a current hot spot. For exam- ensure a high level of robustness and safety of the PLAN
ple, it matches images to a 3D point cloud map (Wolcott system. Tis is because various sensors have diferent
and Eustice 2014). Finally, an example of the 3D-to-3D functionalities. In addition to their primary functional-
matching is the matching of 3D LiDAR measurements ity, each sensor has at least one secondary functionality
and a 3D point cloud map (Wolcott and Eustice 2017). to assist the PLAN of other sensors. Table 10 shows the
According to the prediction algorithm, database- primary and second functionality of diferent sensors in
matching algorithms can be divided into the determinis- terms of PLAN.
tic (e.g., nearest neighbors (Lim et al. 2006) and Iterative Due to their various functionalities, diferent sensors
Closest Point (ICP) (Chetverikov et al. 2002)) and sto- provide diferent human-like senses. Table 11 lists PLAN
chastic (e.g., Gaussian distribution (Haeberlen et al. sensors corresponding to diferent senses of the human
2004), Normal Distribution Transform (NDT) (Biber and body. Te same type of human-like sensors can provide a
Straßer 2003), histogram (Rusu et al. 2008), and machine- backup or augmentation to one another. Meanwhile, the
learning-based) ones. Machine learning methods, such diferent types of human-like sensors are complementary.
as Artifcial Neural Network (ANN) (Li et al. 2019b), Tus, by fusing data from a diversity of sensors, extra
random forests (Guo et al. 2018), Deep Reinforcement robustness and safety can be achieved.

Table 10 Primary and secondary functionality of various sensors in terms of PLAN

Sensor Primary functionality Secondary functionality
HD map Precise localization and environment perception Constrain localization solutions through map matching
LiDAR Precise point‑cloud‑based localization Provide environment models and constraints
Camera Provide visual odometry and visual SLAM, or match Provide environment models and constraints
images with a database
RADAR Ranging and object detection Enhance cameras under challenging illumination conditions
INS Provide continuous self‑contained position, velocity, Bridge outages of other sensors, and aid signal acquisition of other sensors
and attitude
Magnetometer Provide absolute heading Provide position through magnetic matching
Odometer Provide absolute velocity and relative distance Constrain the drift of INS errors, and bridge GNSS and vision signal outages
Pressure Provide absolute height Identify the foor level and constrain the drift of INS altitude errors
GNSS Provide absolute position, velocity and time Help initialization for DR and database matching
UWB Provide absolute position Provide augmentation to GNSS and INS in indoor and urban areas
Ultrasonic Provide absolute position Enhance navigation under challenging weather conditions
Visible light Provide absolute position Reliable landmark updates for DR
WiFi/BLE Provide absolute position Help initialization for INS and database matching
5G Provide absolute position Model multipath environment