stereo-inertial-gnss-lidar device

Fig. 6: Our handheld device for data collection, (a) shows our
minimum system, with a total weight of 2.09 Kg; (b) an additional
D-GPS RTK system and an ArUco marker board [25] are used to
evaluate the system accuracy
R
3LIVE: A Robust, Real-time, RGB-colored, LiDAR-Inertial-Visual tightly-coupled state
Estimation and mapping package

Fig. 1. Falcon 4 UAV platform. Our Falcon 4 platform used for this work is the
successor of our Falcon 450 platform [1] and is equipped with a 3D LiDAR, an
Open Vision Computer (OVC) 3 [2] which has a hardware synchronized IMU
and stereo cameras, an Intel NUC onboard computer, and a Pixhawk 4 flight
controller. The platform has a total weight of 4.2 kg and a 30-minute flight time.

Large-Scale Autonomous Flight With Real-Time
Semantic SLAM Under Dense Forest Canopy

Fig. 3. Multispectral stereo setup with thermal (left) and visible (right)
cameras.

Sky-GVINS: a sky-segmentation aided GNSS-Visual-Inertial system for robust navigation in urban canyons

IC-GVINS: A Robust, Real-time, INS-Centric
GNSS-Visual-Inertial Navigation System for
Wheeled Robot


The proposed IC-GVINS is implemented using C++ under
the framework of Robot Operating System (ROS), which is
suitable for real-time application. The dataset collected by a
wheeled robot is adopted for the evaluation. The equipment
setup of the wheeled robot is showed in Fig. 3. The sensors
include a global shutter camera with the resolution of
1280x1024 (Allied Vision Mako-G131), an industrial-grade
MEMS IMU (ADI ADIS16465), and a dual-antenna GNSSRTK receiver (NovAtel OEM-718D). All the sensors have been
synchronized through hardware trigger to the GNSS time. The
intrinsic and extrinsic parameters of the camera have been well
calibrated using the Kalibr [26] in advance. An on-board
computer (NVIDIA Xavier) is employed to record the multisensor dataset. A navigation-grade [4] GNSS/INS integrated
navigation system is adopted as the ground-truth system. The
average speed of the wheeled robot is about 1.5 m/s.

GVINS: Tightly Coupled GNSS–Visual–Inertial
Fusion for Smooth and Consistent State Estimation

As illustrated in Fig. 10, the device used in our real-world
experiments is a helmet with a VI-Sensor [35] and a u-blox
ZED-F9P GNSS receiver4 attached. The detailed specifications
of each sensor are shown in Table IV. Although the VI-Sensor
provides two cameras as a stereo pair, we only use the left one for
all experiments. The u-blox ZED-F9P is a low-cost multiband
receiver with multiconstellation support. In addition, the ZEDF9P has an internal RTK engine, which is capable of providing
the receiver’s location at an accuracy of 1 cm in an open area.
The real-time RTCM stream from a nearby base station is fed
to the ZED-F9P receiver for the ground truth RTK solution.
In terms of time synchronization, the camera and the IMU are
synchronized by the VI-Sensor, and the local time is aligned with
the global GNSS time via the pulse per second (PPS) signal of
the ZED-F9P and hardware trigger of the VI-Sensor.

A LiDAR-Inertial-Visual SLAM System with Loop Detection

LiDAR-Visual-Inertial Odometry Based on Optimized Visual Point-Line Features

VILO SLAM: Tightly Coupled Binocular Vision–Inertia SLAM Combined with LiDAR

In order to verify the accuracy and effectiveness of the tightly coupled pose estimation algorithm based on binocular VILO proposed in this paper, some extreme experimental environments need to be selected, such as insufficient light or darkness, lack of texture characteristics (indoor white walls), and frequent movement of dynamic obstacles (people). Therefore, this section focuses on the corridor environment of the indoor experimental building with extreme conditions for algorithm comparison and verification experiments. In this paper, we compare the performance of VILO (ours), VINS Fusion and ORB-SLAM2.

In this pose-estimation experiment of the three types of algorithms, it is necessary to use the cumulative error of each algorithm to measure the excellence of each algorithm. Therefore, the visual loop detection function is not turned on during the operation of the three types of algorithms because it would eliminate the cumulative pose error. All the methods are executed on a computing device equipped with an Intel i7-8700 CPU using the robot operating system (ROS) in Ubuntu Linux. The sensor mounting platforms are shown in Figure 4.
Experimental configuration description: The algorithm verification environment is a corridor environment on the first floor of the experimental building with a width of 3 m, a long length, corners, and a relatively empty hall environment. The area on the first floor is about 250 × 100 m2. The scene of this experiment is shown in Figure 5, where (a) is the satellite map of the experimental building, in which you can clearly see the outline of the corridor, and (b–d) show the scene inside the corridor. The robot is controlled to traverse every scene in the corridor as much as possible, the linear velocity of movement is maintained at 0.5 m/s, and the angular velocity is maintained at 0.5 rad/s. A large number of fixed marking points are arranged inside the corridor to obtain the real position of the robot in order to analyze the positioning error of the robot.

A Tightly Coupled LiDAR-Inertial SLAM for Perceptually Degraded Scenes

We conducted a series of quantitative and qualitative analysis experiments on the performance of the proposed tightly coupled LiDAR-IMU fusion SLAM algorithm and compared the results with those of other state-of-the-art LiDAR-SLAM methods. All methods were tested under the same conditions. The hardware platform was an inspection robot with sensors and an on-board computer, as shown in Figure 2. The sensor consisted of a LiDAR (Velodyne VLP-16) with a sampling frequency of 10 Hz and an IMU (hipnuc-CH110) with a sampling frequency of 200 Hz. The on-board computer was an Intel Core i7 with a 2.7 GHz clock, eight cores and 16 G of RAM. All algorithms were implemented in C++ and executed on an Ubuntu 18.04 system using the medoic version of ROS.

An Integrated SLAM Framework for Industrial AMRs with Multiple 2D
LiDAR-Visual-Inertial based on Binary Pose Fusion

4.1. Experimental and Evaluation Setup
In this part, we describe the evaluation results of the proposed SLAM framework. The experimental setup
is demonstrated with two platforms. First, one system is conducted in our laboratory setting to analyze the
performance and robustness. The second system is the double-holonomic robot in manufacturing to confirm the
overall performance.
4.1.1. Experimental Setup
A double-holonomic mobile robot platform with eight mecanum wheels is utilized to evaluate the performance
in manufacturing, as shown in Fig. 2. The sensor system for the robot is equipped with two Sick 2D LiDAR sensors
and two Zed 2 cameras, IMU. We used a Linux-based operating system and ROS Melodic with C++ language on
the embedded computer NVIDIA Jetson AGX Xavier to manage the SLAM system.
In order to analyze the performance and robustness, we also build a sensor setup similar to the actual robot
with two 2D LiDAR sensors, two stereos Zed 2 cameras and an Xsense MTi-10 IMU, as shown in Fig. 9. Table 1
shows the characteristics of sensors used in the proposed method, which is not used high-performance sensors like
in the actual holonomic robot.
We implement Algorithm 2 following the pipeline, as described in Fig. 3. The SLAM system is handled with
multiple parallel threads for synchronization, laser merging and filtering, state estimation, and SLAM solver.

Stereo Visual Inertial LiDAR Simultaneous Localization and Mapping

A. Platform and software
We built a platform (Fig. 1(a)) with two megapixel
cameras, a 16 scan-line LiDAR, an IMU (400Hz), and
a 4GHz computer (with 4 physical cores). We built a
custom microcontroller based time synchronization circuit
that synchronizes the cameras, LiDAR, IMU and computer
by simulating GPS time signals. The software pipeline is
implemented in C++ with ROS communication interface. We
use GTSAM library [40] to build the fixed-lag smoother
in the VIO. For loop closure, we use ICP module from
LibPointMatcher [41] to align point clouds, DBoW3 [42] to
build the visual dictionary, and iSAM2 [36] implementation
in GTSAM [40] to conduct global optimization.

On Visual-Aided LiDAR-Inertial
Odometry System in Challenging
Subterranean Environments
Hengrui Zhang
CMU-RI-TR-21-35
August 11, 2021
https://www.ri.cmu.edu/app/uploads/2021/08/Henry_Thesis_Final.pdf


In the DS Drone setup, we choose IMU internal clock time as our time server,
and the other sensors synchronize their timestamps to the IMU time server.
To synchronize the Velodyne to the IMU clock, the PPS signal from the IMU
(based on its internal clock) is directed to the Velodyne. From the NUC computer
side, when it receives a PPS IMU data packet, it creates a fake NMEA GPS message
with a timestamp rounded to closest integral second and sends the NMEA message
to the Velodyne through an Ethernet connection.
For the uEye camera, the system uses a Teensy 3.2 microcontroller to manage the
camera external triggering signal. The IMU PPS signal is wired to a GPIO pin on
the Teensy so it has IMU internal clock information available. Upon image request,
the uEye driver sends the start camera trigger command to the Teensy, and then
Teensy generates a 15 hz camera trigger signal starting from the next IMU PPS.
This ensures that the camera grabs frame at 15 hz, and the first image aligns with
an integer second in the IMU time frame. The camera internal clock timestamp of
the first image is also stored, and used as the base stamp to calculate incremental
time differences for the following images (as shown in figure 3.3). Note that this
scheme assumes that the total flight time of the drone is short enough that the camera
internal clock and IMU internal clock (server) do not drift too much. We do confirm
this is the case for DS drones with its 15 minutes flight time.
The above synchronization scheme avoids modifying the system clock, which takes
time when syncing through tools like Chrony, and potentially disrupts other processes
that rely on computer system time. This is especially important when launching
drones autonomously in the field in a short period of time.
3.3 Image Quality
Image quality is crucial for visual odometry systems. To get the best out of our
camera sensor, we did some custom tuning to the image quality. In this section,
we will present the camera model, exposure compensation, and image brightness
adjustment to achieve good image quality.
Camera Sensor and Lens Model
The main camera on DS drones is the uEye UI-3271LE-C-HQ with a Sony IMX265
imaging sensor. It has a 1/1.8” global shutter sensor as shown in figure 3.4. Comparing
with its Intel RealSense [11] counterparts on the ground vehicles, the global shuttering
sensor reduces the skew of objects in images under fast motion. Without an online
compensation for rolling shutter effects, global shutter cameras are considered to be
the better choice for state estimation purposes.
We choose the Lensagon BF5M2023S23C fisheye lens for the camera. The wideangle fisheye lens gives us 195◦ horizontal and vertical field of view. As shown in the
figure 3.5, due to the limited sensor size, some cutoff of FOV appears in the vertical
direction, but the overall FOV satisfies the requirements and gives us a better view
of the scene.
To model the lens for VIO purposes, we adopt to use a pinhole camera model
with equidistant lens distortion model [10]. The equidistant distortion can model our
lens distortion well. An undistorted image is shown in figure 3.5.

Super Odometry: IMU-centric LiDAR-Visual-Inertial Estimator for
Challenging Environments

We collected our test dataset with Team Explorer’s DS
drones (Fig.4(e)), deployed in the DARPA Subterranean
Challenge. It has a multi-sensor setup including a Velodyne
VLP-16 LiDAR, an Xsens IMU, a uEye camera with a wideangle fisheye lens, and an Intel NUC onboard computer.
The data sequences were designed to include both visually
and geometrically degraded scenarios, which are particularly
troublesome for camera- and LiDAR-based state-estimation.

LVI-SAM
This repository contains code for a lidar-visual-inertial odometry and mapping system, which combines the advantages of LIO-SAM and Vins-Mono at a system level.

The datasets used in the paper can be downloaded from Google Drive. The data-gathering sensor suite includes: Velodyne VLP-16 lidar, FLIR BFS-U3-04S2M-CS camera, MicroStrain 3DM-GX5-25 IMU, and Reach RS+ GPS.

https://drive.google.com/drive/folders/1q2NZnsgNmezFemoxhHnrDnp1JV_bqrgV?usp=sharing
Note that the images in the provided bag files are in compressed format. So a decompression command is added at the last line of launch/module_sam.launch. If your own bag records the raw image data, please comment this line out.

LiDAR-Visual-Inertial Odometry Based on Optimized Visual
Point-Line Features

To evaluate the performance of the algorithm we conducted in the outdoor environment, the Hong Kong dataset was used for performance evaluation and it was compared
with other similar advanced algorithms. The experimental equipment and environment
are shown in Figure 14. The sensor models are as follows: the camera is BFLY-U3-23S6CC, the LiDAR is HDL 32E Velodyne, IMU is Xsens Mti 10, and the GNSS receiver is u-blox
M8T. In addition, we utilized the high-grade RTK GNSS/INS integrated navigation system, NovAtel SPAN-CPT, as the ground truth