IEEE Robotics & Automation Magazine - September 2023 - 32

Two mechanical lidar sensors (i.e., VLP-16 and RPLIDARA2)
were able to detect objects more than 7 m away, whereas
solid-state lidar (i.e., CE30-D and HPS-3D160) could not
detect objects farther than 4 m away. Overall, the number of
point clouds acquired by the lidar sensor gradually decreased
as the distance from the rods increased. In addition, lidar
detection performance was influenced by the material and
shape of the rods. In particular, in the case of steel and acrylic,
the measurements were inaccurate due to the scattered reflection
and absorption, depending on the distance. According to
the experimental results, CE30-D was not able to detect the
acrylic rods at any distance, and HPS-3D160 only detected
them at 2 m. In conclusion, the degree of accuracy of lidar
sensors in detecting objects decreased in the order of wood,
steel, and acrylic.
For radar sensors, the range profile for the distance of 2 m
is plotted in the last two rows of Figure 3. Both radars clearly
recognized every object up to 7 m, and the peak amplitude
gradually decreased with distance. The steel and wood detection
results show the highest and the lowest amplitude peaks,
respectively. Because all conditions were identical except
for the material of the object, it may be assumed that detection
of the material was better at higher peak amplitudes. In
conclusion, the degree of detection accuracy in radar sensors
decreased in the order of steel, acrylic, and wood.
Figure 4(b) shows the point cloud data from four type of
lidar sensors. As the four lidars possess different FoV and sensing
mechanisms, the point clouds obtained from each sensor
were distinct. To clearly visualize individual measurements,
the objects were schematically drawn and the FoV of each sensor
was indicated in gray [see Figure 4(c)]. The mechanical
lidars, including VLP-16 and RPLIDAR-A2, exhibit a complete
360° horizontal FoV, and their point data are clear as the
laser intensity is relatively high. However, as RPLIDAR-A2 is
for 2D scanning, it suffers from detecting objects that the 2D
laser plane penetrates (e.g., a chair). On the other hand, static
lidars, including CE30-D or HPS-3D160-U, possess a narrow
FoV compared to mechanical lidars and yield scattered point
cloud data. However, they exhibit an increasing tendency to
recognize transparent objects such as acrylic boxes. The detection
ability of an FMCW radar sensor is graphically drawn as
2D mapped normalized amplitudes [see Figure 4(d)]. Although
exact geometry information of the objects is not obtained, perception
of the object by FMCW radar is observed to be less
affected by their material or shape than by lidar sensors.
TABLE 2. Specifications of radar sensors.
SENSORS MECHANISM
IWR 1443
[26]
X4M03
[27]
Pulsed
wave
NUMBER OF
RECEIVERS
CW FMCW 4
1
NUMBER OF
TRANSMITTERS
3
1
TRANSMITTER
POWER
(DBM)
12
Low: 0.7;
medium: 4.1;
high: 6.3
32 IEEE ROBOTICS & AUTOMATION MAGAZINE SEPTEMBER 2023
FREQUENCY
76-81 GHz
7.29-
8.75 GHz
COST
(US$)
BANDWIDTH
(GHZ)
299
4
425 1.5
OPERATING
VOLTAGE (V)
3.3/1.8
BODY SIZE
(MM × MM)
104 × 104
Maximum 3.6 58 × 30
MAPPING
SETUP
Mapping is one of the most important application fields for realizing
autonomous mobile robots. The quality of maps developed by
different sensors have been an important consideration for selecting
the appropriate sensor for a target application. We therefore
examined the accuracy of map generation from the scanned data
obtained from lidar sensors installed on a mobile robot. As radar
sensors possess low resolutions and low-density point clouds, they
are not suitable for solitarily being used in environmental reconstruction.
Therefore, only lidar sensors were considered for mapping.
Four different types of lidar sensors were placed on top of a
mobile robot, Pioneer 3-AT, to scan the environment during
motion. An embedded computer board (a Jetson AGX Xavier
Developer Kit) was used to execute the gMapping algorithm
(http://wiki.ros.org/gmapping) under the Robot Operating Systemmelodic
software platform on Ubuntu 18.04. The Mapping algorithm
considers as inputs the odometry of the mobile robot, which
is obtained from wheel encoders, and the consecutive 2D lidar
scanned data. In the case of 3D lidar, 2D lidar scanned data are
obtained from the projection to 2D data. For an unbiased comparison,
conditions such as odometry uncertainty and calibration
parameters were applied equally to all the employed lidar sensors.
The ground-truth map is shown in Figure 5(a), and the mobile
robot moves along the path in the area marked in orange.
RESULT
To verify the accuracy of the maps generated by each sensor,
they were overlaid on the actual 2D drawing and compared, as
depicted in Figure 5(b)-(d). It can be seen that two mechanical
lidar sensors generate a high-accuracy map because of a full
FoV of 360°. In particular, VLP-16 can scan a wider range
than RPLIDAR; thus, the resulting map is more sophisticated
and covers a larger region for the same route. Meanwhile, solid-state
lidar sensors generate less accurate maps because of
their narrow FoV of 60° for CE30-D and 72° for HPS-3D160.
The accuracy of the maps generated by CE30-D is inferior
over a large area, and HPS-3D160 is unable to generate the
map completely because of its markedly insufficient number
of points in a point cloud. The experimental results show that
FoV of the sensors was crucial for mapping accuracy in
wide spaces, and detection range of the sensors exerts an
effect on the coverage of the map generated under a constant
moving trajectory. Specially, the narrow FoV of CE30-D and
http://wiki.ros.org/gmapping
IEEE Robotics & Automation Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2023
