Page 15 - 卫星导航2021年第1-2合期

El‑Sheimy and Li Satell Navig (2021) 2:7 Page 5 of 23

Table 4 Companies and products in indoor PLAN
Primary sensor Company Performance Sensor cost
Camera Mobileye Decimeter‑level to centimeter‑level $100 level to $10 level
Camera (infrastructural) VICON Centimeter‑level and higher $10,000 level to $1,000 level for a specifc area (e.g., a park‑
ing lot)
LiDAR Velodyne Decimeter‑level to centimeter‑level $10,000 level to $1,000 level
LiDAR (infrastructural) SICK Decimeter‑level to centimeter‑level $10,000 level for a specifc area
HD map HERE Centimeter‑level Costly. A team of professional cars works for days to create
a regional HD map
Inertial sensors (vehicle) Profound positioning 2% of travel distance Low‑cost IMU at $ 100 to $ 10 level
Inertial sensors (pedestrian) TDK‑InvenSense 4–8% of distance; 1% of distance Low‑cost IMU at $ 100 to $ 10 level
with a fxed device
GNSS Trimble Decimeter‑level to centimeter‑level $ 1,000 level
UWB Decawave Decimeter‑level $ 100 level for a set of anchors that are used to cover a
2
place of 100 m level
Ultrasonic Marvelmind Decimeter‑level $ 100 level for a set of anchors that are used to cover a
2
place of 100 m level
WiFi/BLE Cisco 50% within 5 m and 90% within 10 m Infrastructure deployment cost of $ 10 level per 100
2
m ‑level area
5G Huawei No commercial PLAN system yet $1000 to $ 100 level per base station. Coverage range from
kilometer‑level to within 100 m

In the past decade, HD maps have received extensive Furthermore, to update the HD map in an area efectively
attention. An important reason is that traditional maps where changes have occurred, there are challenges in
are designed for people, not machines. Terefore, the transmitting, organizing, and processing massive crowd-
accuracy of the traditional map cannot meet the require- sourced data. For example, one hour of autonomous driv-
ments of autonomous driving. Also, the traditional ing may collect one terabyte of data (Seif and Hu 2016). It
map does not contain enough real-time information for takes 230 days to transfer one week’s autonomous driving
autonomous driving, which requires not only informa- data using WiFi (MachineDesign 2020). Tus, dedicated
tion about the vehicle, but also information about exter- onboard computing chips, high-efciency commu-
nal facilities (Seif and Hu 2016). With these features, the nication, and edge computing are needed. Terefore,
HD map is not only a map but also a "sensor" for PLAN crowdsourcing HD maps requires cooperation from car
and environment perception. Table 5 compares the tradi- manufacturers, map manufacturers, 5G manufacturers,
tional map and HD map. and terminal manufacturers (Abuelsamid 2017).
HD map is key to autonomous driving. It is generally
accepted that HD maps require centimeter-level accuracy
and ultra-high (centimeter-level or higher) resolution. LiDAR
Accordingly, creating HD maps is a challenge. Te crea- LiDAR systems use laser light waves to measure distances
tion and updating of the current HD maps are depend- and generate point clouds (i.e., a set of 3D points). Te
ent on professional vehicles equipped with high-end distance is computed by measuring the time of fight of
LiDAR, cameras, RADARs, GNSS, and INS. For example, a light pulse, while the direction of a transmitted laser is
Baidu spent 5 days building an HD map in a Beijing park tracked by gyros. By matching the measured point cloud
by using million-dollar-level mapping vehicles (Synced with that stored in a database, an object can be located.
2018). Such a generation method is costly; also, it is dif- LiDAR is an important PLAN sensor on unmanned
fcult to update an HD map continuously. vehicles and robots. Figure 2 compares the PLAN-related
To mitigate the updating issue, crowdsourcing based performance of the camera, LiDAR, and RADAR.
on car-mounted cameras has been researched. Tis Te main advantages are its high accuracy and data
method can lower the requirement of extra data collec- density. For example, the Velodyne HDL-64E LiDAR has
tion if the images from millions of cars are used properly. a measurement range of over 120 m, with ranging accu-
However, this task is extremely challenging. First, it is racy of 1.5 cm (1 sigma) (Glennie and Lichti 2010). Te
difcult to obtain the PLAN solutions that are accurate observation can cover 360° horizontally, with up to 2.2
enough for HD map updating with crowdsource data. million points per second (Velodyne 2020). Such features