IEEE Robotics & Automation Magazine - March 2021 - 52

sensors or human intervention. We assumed that three
linearly independent planar surfaces forming a wall corner shape are provided as the calibration targets, ensuring that the geometric constraints are sufficient to
calibrate each pair of lidars. After matching the corresponding planar surfaces, our method successfully recovered the unknown extrinsic parameters with two steps: a
closed-form solution for initialization based on the Kabsch
algorithm [6] and a plane-to-plane iterative closest point
for refinement.
The overview of our 3D object detection is depicted in
Figure 4. The inputs to our approach are multiple point
clouds captured by different lidars. We adopted an early
fusion scheme to fuse the data from multiple calibrated
lidars at the input stage. With the assumption that the
lidars are synchronized, we transformed the raw point
clouds captured by all of the lidars into the base frame
and then fed the fused point clouds into the 3D object
detector [4]. The final output was a series of 3D bounding boxes.

3D Boxes

VoxelNet

3D Object Detector

Multi-Lidar Fusion

Multi-Lidar Input

3D Point-Cloud Mapping
3D point-cloud mapping aims to build a 3D map of the
traversed environments. Figure 5 illustrates the diagram
of our mapping system. The inputs to the system are the
raw data from the IMU and 3D lidar (i.e., accelerometer
and gyroscope readings from the IMU and point clouds
from the 3D lidar). The system starts with an adapted
initialization procedure followed by two major submodules: lidar inertial odometry and rotationally constrained mapping. Since the vehicle usually remains still
at the beginning of mapping, we do not need to excite
the IMU to initialize the module as described in [7],
which is more suitable for handheld applications. With
the stationary IMU readings, the initial orientation for
the first body frame can be obtained by aligning the
average of the IMU accelerations to the opposite of the
gravity direction in the world frame. The initial velocity
and IMU biases are set to zero. Then, the lidar inertial
odometry optimally fuses the lidar and IMU measurements in a local window. The mapping with rotational

Figure 4. An overview of the 3D object detection module. The inputs are multiple point clouds captured by synchronized and wellcalibrated lidars. We used an early fusion scheme to fuse the data from multiple calibrated lidars and adopted the VoxelNet [4] to
detect 3D objects from the fusion results.

3D Lidar Point Cloud
Undistorted
Point Cloud

Feature Extraction
and Local Map
Management

Propagation and
Preintegration

Joint Nonlinear
Optimization

Refined Global Map
and Lidar Pose

Initialization

IMU Accelerometer
and Gyroscope Readings

Lidar-Inertial Odometry

Undistorted
Point Cloud
and Lidar Pose

Rotationally
Constrained
Mapping

Figure 5. The schematic diagram of our 3D point-cloud mapping system. After the initialization, the system estimates the states and
refines the global map and lidar poses in, respectively. The odometry and mapping submodules.

IEEE ROBOTICS & AUTOMATION MAGAZINE

MARCH 2021

IEEE Robotics & Automation Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - March 2021