IEEE Robotics & Automation Magazine - September 2023 - 37

was 2 m away from a radar sensor. The range profiles of the two
radar sensors were almost constant with respect to time. Both
radar sensors were robust to the fog particles.
To quantitatively evaluate the performance degradation of
sensors under foggy environments, we have detected distances
and counted the number of points for 10 min with fog applied,
in the case of radar and lidar sensors, respectively. For radar
sensors, the distance is detected at the point where the received
power is maximum. All measured distance data and the number
of points are averaged to obtain their means, and then the corresponding
standard deviations (std devs) can be computed as well.
From these means and std devs, the corresponding ratios of the
former to the latter are listed in Table 3, where it can be observed
that the two employed radar sensors were less susceptible to fog.
The smaller the ratio, the better. More intuitively, it can be said
that the smaller the fluctuation of the sensing distance values
due to fog, the stronger the detection ability. In this article, the
aforementioned ratio criterion is called the normalized std dev
(std dev/average) to be consistent with that of lidar sensors.
ODOMETRY-BASED SCANNING WHILE MOVING
Here we consider the mechanism of scanning while in motion.
As in the previous experimental setting, fog was generated in an
indoor space using a fog machine (Rex-10). To confirm fog density
quantitatively, three visible signs were installed 2.5, 5, and
7.5 m away from the sensors, respectively. For comparison, as
fog concentration increased, experiments were conducted at the
moment when visibility was 5, 2.5, and less than 2.5 m. Scanning
was performed with an AgileX Scout minimobile robot
performing two laps along a rectangular path. An inertial measurement
unit (IMU) and 3D mechanical lidar (Ouster OS1)
sensor were installed on top of the mobile platform, and the trajectories
were calculated and compared with those in a fog-free
environment. The employed lidar-inertial odometry (LIO) algorithm
is a revised version of the cartographer (https://googlecartographer-ros.readthedocs.io/en/latest)
algorithm that can be
appropriately used with 3D lidar. To validate the individual performance
of lidars, influence of the IMU sensor was minimized
by deliberately using extremely small weights on the
IMU measurements during the filtering algorithm computation.
The LIO algorithm uses a coarse-to-fine strategy for scan
matching that aligns the current scan of an environment with an
already-built map. For an unbiased comparison, the same algorithm
was applied to all the experimental settings.
Figure 7 shows the trajectories computed using the LIO
algorithm. To obtain measurements from the lidar sensor, the
mobile robot completed two laps along a rectangular reference
trajectory. The tracking performance was evaluated for three
different fog-density levels. As the fog density increased over
time, the laser was more scattered, resulting in increased data
loss and scan-matching failure. Finally, the estimated trajectory
drifted away from the reference. It was observed that, at a
reduced visibility of less than 2.5 m, more than 95% of the data
was lost, causing failure of LIO estimation. Minor data distortion
and loss at low fog density may be overcome by applying
additional data preprocessing, such as outlier removal [18].
GLASSED-IN ENVIRONMENT
As seen in the " Object Detection " section, lidar functions
poorly when attempting to detect objects in a field containing
transparent materials such as glass or acrylic. To demonstrate
the effect of glass on lidar, we evaluated the LIO algorithm in a
space with a glass wall on one side. The glass wall was resized
by blinding it, and three different environmental settings were
prepared: fully blinded, half blinded, and not blinded, as illustrated
in the first row of Figure 8. Similar to the experiments in
a foggy environment, a mobile robot with an Ouster OS1 lidar
sensor moved along a rectangular path and the corresponding
odometry was computed by matching the lidar scan data to the
already-built map point cloud. The LIO algorithm employed
here was the same as that which was used for the experiments
in a foggy environment.
It is observed in Figure 8 that the proximity of the large glass
walls resulted in distortion in the LIO. This distortion was worsened
by the large incident angles of the emitted beams falling on
the glass walls. Specifically, if a lidar sensor is near glass walls, its
laser beams are reflected off them and then returned through two
or more paths due to large incident angles. Such undesirable multipath
interference leads to the inaccurate estimate that the glass
walls are farther away than they actually are, which causes lidar
sensors to match the far point cloud with existing maps. Finally,
the LIO tended to be smaller in the axial direction where the glass
wall was present. Similarly, incorrect maps are generated because
of glass reflection. To compensate for these drawbacks, additional
algorithms are required to classify the true point cloud and reflected
point cloud before the scan-matching step [19]. In addition,
detecting transparent objects using a multiecho laser scanner may
be another potential alternative [20].
DISCUSSION
LIDAR
A mechanical lidar sensor possesses a wide horizontal FoV
and high resolution due to its rotating mechanism. This renders
it suitable for applications that require consecutive matching of
scanned data, such as the mapping or odometry estimation as
confirmed in the " Mapping " section. Furthermore, they possess
a high angular and distance resolution, detecting even thin
rods in the experiments discussed in the " Object Detection "
section. Additionally, they provide relatively robust measurements
TABLE 3. The normalized standard deviation (std dev) of fog environment sensing tests.
SENSORS
Normalized std dev/average
VELODYNE VLP-16
0.26
RPLIDAR-A2
0.86
CE30-D
5.7
HPS-3D160-U
7.66
IWR 1443
0.0044
X4M03
0.0025
SEPTEMBER 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
37

https://google-cartographer-ros.readthedocs.io/en/latest https://google-cartographer-ros.readthedocs.io/en/latest

IEEE Robotics & Automation Magazine - September 2023

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