IEEE Robotics & Automation Magazine - June 2014 - 35

southern Selat Puah marine environment. Data from the
radar, global positioning system (GPS), and an inexpensive
single-axis gyro were logged using an on-board processing
unit as the ASC traversed the enviromnent, which comprised
geographical and sea-surface vessel landmarks. The standard, automated OS-CFAR feature detector, introduced in
the "Feature Detection with Radar" section, was applied to
the radar data to provide the features to be input to the
SLAM algorithms. With restrictive feature modeling and a
lack of vehicle control input information, it is demonstrated
that by adopting the RFS concepts and the PHD filter, useful
localization and mapping results can be obtained, despite an
actively rolling and pitching ASC on the sea surface. The vector-based SLAM algorithm MH-FastSLAM is also implemented and compared.
The ASC and the Coastal Environment
The ASC was originally developed at the Department of
Mechanical and Ocean Engineering, Massachusetts Institute of
Technology (MIT), for experiments in autonomous navigation
in rivers and coastal environments [16]. For stabilization, lateral
buoyancy aids were added to the platform, as depicted in Figure 6. The figure shows the ASC at sea. with the X-Band radar
mounted on a 1.5-m-length pole above the sea surface. The
X-Band radar used was the M-1832 BlackBox Radar from
Furuno and was primarily used to detect buoys and ships at
large distances (several kilometers), which were approximated
to be point features. The mechanically scanned beam has a
width of 3.9° in azimuth and 20° in elevation. The large elevation beam width makes the sensor robust to the sometimes
severe pitch and roll of the ASC. A GPS receiver (Crescent
Hemisphere 110) as well as a KVH Industries, Inc., DSP5000
single-axis gyroscope for 3-D pose (x k, y k, z k) measurements
were also used in the experiments. An on-board processing
unit logged the GPS and gyro data at a rate of 1 Hz, with the
radar data being sampled and logged at a scan rate of 0.5 Hz,
i.e., one full 360° sweep of the environment required 2 s. The
radar-range bin resolution, dr (q), was set to 7.5 m, with a maximum range of 7.68 km. All power values that exceeded the
OS-CFAR threshold were considered as valid point features in
the RB-PHD-SLAM and MH-FastSLAM experiments.
Together with the known GPS locations of the surrounding buoys, an automatic identification system (AIS) receiver
was used for ground truth verification of the map features in
the experiments. An AIS is a short-range coastal tracking
system used for identifying and locating sea vessels by electronically exchanging data. This enables the system to receive
position and speed estimates from a large number of vessels
present in the area. Since those vessels were used as features
in the SLAM algorithms, this source of information was used
as ground truth to verify and compare the features extracted
from the radar data, with the position delivered by the AIS.
The ASC Process Model
A sea-based ASC is subject to numerous uncertain disturbances such as currents and wind, moving the ASC in any

Figure 6. The ASC-adapted kayak.

arbitrary direction. To account for this, the following nonlinear process model f veh ($) in (3) is adopted:
x k = x k - 1 + Vk - 1 3 Tk cos ^z k - 1 + dz k - 1h + v kx - 1
y
y k = y k - 1 + Vk - 1 3 Tk sin ^z k - 1 + dz k - 1h + v k - 1

= z k - 1 + dz k - 1 + v k - 1,
i.e., X k = f veh ^ X k - 1, U k - 1, v k - 1h,
z

(11)

where x k, y k, and z k represent the easting, northing, and ASC
heading angle with respect to north at time k, X k =
[x k y k z k] T and f veh ($) is the vehicle motion vector function
encapsulating (11). U k -1 represents a vector comprising the
input velocity signal and the measured angular change, i.e.,
U k -1 = [Vk -1 dz k -1] T , recorded by an on board single axis
y
z
gyroscope. Here, v xk -1, v k -1 and v k -1 represent random perturbations in the ASC motion due to external sea forces and
are modeled by white Gaussian signals, encapsulated in the
y
z
noise vector v k -1 =[v xk -1 v k -1 v k -1] T . Here, DTk =t k - t k -1
is determined from the
measurement rate of the
gyro. In this experiment,
The RB-PHD-SLAM
for simplicity, Vk = Vk -1
and is chosen a priori due
approach can be seen to
to the lack of suitable
Doppler velocity log sengenerate more accurate
sors. A constant-velocity
model could also be
localization and feature
assumed, accompanied by
the recursive estimation of
number estimates.
Vk integrated into the
SLAM algorithm. This
vehicle process model will be used in both the SLAM algorithms, developed for comparison purposes, in this article.
Vector-Based Multihypothesis
FastSLAM Comparison
FastSLAM estimates the map on a per-particle basis, meaning
that different particles can be associated with different features
[12]. This means that the FastSLAM filter has the possibility to
June 2014

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2014