IEEE Robotics & Automation Magazine - June 2020 - 140

Even though self-driving technologies have been studied
for a few decades [1] and several successful applications have
been demonstrated in urban areas [2], rural roads [3], and
agricultural environments [4], autonomous industry vehicles
are currently applied only in some specific scenarios with very
low speed and fixed routes and are typically implemented in
an indoor closed area of small-scale warehouses [5]. Compared to on-road self-driving cars, autonomous industry vehicles have several important differences. First, in many tasks,
the localization and navigation accuracy must be as high as
20-50 mm, which is much higher than the lane-level accuracy
required for typical on-road self-driving cars. Second, the
industrial environment is semiclosed but can be very large
scale, highly dynamic, and feature sparse, which brings great
difficulties to place-recognition, loop-closure, and mapping
tasks. Third, vehicles may need to move in more unorganized
environments. To meet stringent safety requirements, the sensing system should be more accurate and robust. These differences present new challenges in developing the mapping,
localization, and perception systems of industrial vehicles.
In this article, we introduce the main perception and localization technologies of the proposed autonomous transportation vehicles used to achieve full autonomy in cargo
transportation in large-scale airport terminals and tarmac environments. The main contribution of the article is to present a
system-level overview of the deep learning-enhanced, largescale localization and mapping approaches as well as multisensor fusion-based perception algorithms. We also demonstrate
how these technologies can enable autonomous cargo transportation in large-scale airport environments.
Application Scenario and Related Work
As shown in Figure 1, in this article, we consider the industrial vehicle that delivers cargoes and materials in airport,
harbor, and automated warehouse environments. Typically,
these environments are semistructured and unorganized
and have a very large-scale area with both indoor and outdoor routes. While GPS and inertial navigation systems are
widely used in autonomous vehicles to present real-time
location information, they still cannot fulfill the autonomous navigation requirement, especially in industrial applications. Firstly, the positioning error of the sensor system
mentioned previously cannot, in general, fulfill the localization accuracy required by the high-precision navigation system in cargo transportation vehicles. Secondly, GPS often
does not work because of the shelter of houses and overpasses and also suffers from various signal interference in
industrial environments. Another important issue is the
sparse static structured information. Since airport, harbor,
and warehouse environments usually contain a number of
dynamic objects (such as airplanes and vehicles) and temporarily stored entities (such as containers, pallets of goods,
and industrial materials), autonomous navigation in such a
highly dynamic environment requires higher localization
accuracy than that of the current GPS-based inertial guidance systems for household vehicles. These factors bring
140

IEEE ROBOTICS & AUTOMATION MAGAZINE

JUNE 2020

great difficulties to the construction of the environment
map and estimation of the vehicle pose.
In recent years, simultaneous localization and mapping
(SLAM) is increasingly promoting the development of the
fully autonomous navigation of mobile robots and vehicles. A
typical SLAM system estimates the real-time vehicle pose and
builds/updates the environment map simultaneously. In practice, for a semiclosed industry scenario, building a complete
map beforehand is recognized as a more effective method due
to its strong robustness to dynamic objects, high precision in
the presence of off-line refinements, and better localization
performance over a long period of time.
With a prior environment map, autonomous vehicles have
the ability to plan global routes and local paths, estimate their
locations with onboard sensor information, and achieve
autonomous navigation. Current solutions include vision [6]
and laser point cloud-based approaches [7]. Vision-based
solutions are unreliable in industrial environments due to the
limited field of view, motion blur, and nonrobustness to variable weather and illumination conditions. Advanced solutions
based on laser point cloud are more popular in practical applications. A typical solution is to extract handcraft features
(ground, edge, and planar points) from laser data and estimate
the vehicle pose with odometry data and feature-matching
results. Then parallel localization and mapping can be
achieved by implementing the Iterative Closest Point (ICP)based graph-optimization framework [7]. Recently, to
enhance the generalization ability of extracted laser features
and improve algorithm robustness, deep learning-based
approaches have become more and more popular [2]. However, in large-scale, dynamic environments with sparse static features and large cumulative errors, mapping and localization
tasks are still very challenging [8]. In this article, we present a
point cloud description network to extract discriminative and
generalizable laser features and achieve large-scale place recognition tasks. Based on this, we developed a deep learningbased integrated framework for mapping and localization
tasks in large-scale dynamic environments.
In industrial environments, autonomous vehicles need to
share the drivable area with human workers, forklifts, and
other vehicles. The driving environment is usually crowded
and narrow. Safety concern is the first priority, especially in airport and harbor environments. Vehicles need to be able to
detect surrounding dynamic entities and track their poses.
These help the autonomous vehicle have a deeper understanding of the surrounding scenario, predict short-term future
information, and make sophisticated precedence decisions.
For example, a detection and tracking system that estimates
the state of surrounding pedestrians and vehicles provides the
motion control system the capability of avoiding obstacles,
thus ensuring safety. Computer vision-based approaches
incorporated with deep learning techniques have been widely
studied and have achieved remarkable performance in object
detection and tracking tasks [15]. However, the illumination
condition of industrial environments changes wildly, so we
developed a multimodel information-fusion-based perception

IEEE Robotics & Automation Magazine - June 2020

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