IEEE Robotics & Automation Magazine - June 2020 - 143

Mapping

JUNE 2020

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

Figure 4. The system structure of the proposed deep learning-based, large-scale localization and mapping approach.

Place Descriptor Mapping
Environment Analysis

Typical Place
Analysis

Place Descriptor Map

Local Sequence
Matching
Place Description Network
(Descriptor Generation)
Local Feature Extraction

3D Laser
Point Cloud

Large-Scale Point Cloud Learning

Coarse Matching

Loop Closure

Point Cloud
Registration

Pose-Graph Optimization
Feature Extraction and Matching

Ego-Motion Estimation

Key Frame Generation

Point Cloud Mapping
Pose Prediction
Odometry

Pose Tracking

localization, it is necessary to build a prior
map of the whole environment, and the
mapping accuracy limits the errors
incurred in estimating the pose of the
vehicle. In this article, we present a deep
learning-enhanced mapping and localization framework for large-scale industry
environments. As shown in Figure 4, the
proposed framework includes the following components:
● Large-scale point cloud learning: We first
extract local features from the raw 3D
laser data to reveal the local 3D structure around each point. An entropy minimization-based criterion is utilized to
determine the neighborhood size, and
three kinds of local features are calculated: 2D/3D eigenvalue-, height distribution-, and normal vector-based features.
Using these local features as inputs, we
then develop a deep learning-based
place description network to generate a
discriminative and generalizable global
descriptor to uniquely encode the largescale point cloud. More specifically, the
graph neural network is introduced to
reveal the intrinsic relations among different compositions in the point cloud
and aggregate the spatial distribution
characteristics of similar local structures; this contributes to aggregating the
sparse static structured information in
the large-scale environment and improve
robustness to dynamic objects. More
details of the proposed place description
network can be found in [10].
● Place descriptor mapping and environment analysis: The generated descriptors with their associated positions
compose a place descriptor map. Then
we cluster the descriptors in the feature
space (based on the L 2 distance) to find
out the places corresponding to cluster
centers, and we define these places as
typical places.
● Loop closure: We introduce a coarse-tofine strategy to detect loop-closure candidates. More specifically, after generating a
new place descriptor, we first compare it
with all the typical places to roughly estimate its location and then conduct a
series of local place matchings around
the typical place to find out the accurate
location. A sequence-matching strategy is
introduced to exclude interference
brought by extremely similar places.

*

143
IEEE Robotics & Automation Magazine - June 2020

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2020
