17-06-2013, 03:01 PM
Heuristics-enhanced Odometry for Indoor Tracking of Segways
Heuristics-enhanced.pdf (Size: 1.15 MB / Downloads: 22)
ABSTRACT
This White Paper introduces a method for dramatically improving the accuracy of odometry
for vehicles driven in structured indoor environments. The method is suitable for all vehicles
controlled by human operators, such as the Segway HT or PT. In this paper, we refer to all those
human-operated vehicles collectively as “Personal Vehicles” (PVs). Our proposed method,
called “Heuristics-Enhanced Odometry” (HEO) can also be applied to some, but not all
autonomous mobile robots.
HEO uses heuristic assumptions about the environment to correct odometry errors. The two
main assumptions are that most indoor travel happens along straight corridors, and that most
corridors intersect at 90° angles (45° angles are also acceptable). When these assumptions are
true, HEO produces near-zero heading errors for runs of unlimited duration. This performance is
sufficiently accurate to track the position of PVs for extended periods of time. HEO requires no
hardware or sensors other than two wheel encoders. The algorithm can be implemented on
simple microprocessors and with as little as ~20 lines of program code.
INTRODUCTION
Arguably the most widely used method for tracking the position of vehicles is odometry. In
computer-controlled vehicles, such as mobile robots or remotely operated ground vehicles,
optical incremental encoders are mounted on at least one left and one right wheel to count
fractions of wheel revolutions. The difference between left and right encoder pulse counts is
directly proportional to changes in heading, and the average of left and right encoder pulse
counts is directly proportional to linear displacement. With the well known equations of
odometry it is very easy to compute the relative change of position of a vehicle, making
odometry an inexpensive solution for position estimation.
SEGWAY HT ODOMETRY SYSTEM
For the implementation and testing of the HEO
algorithm on a human-controlled PV, we chose a
Segway Human Transporter (HT), shown in
Figure 1. The Segway HT does not allow useraccess
to odometry data. For this reason we
designed and implemented our own quadrature
encoders.
Since the Segway’s drive motors are not
easily accessible, we decided to mount encoders
directly to the wheels, which can be removed
from the vehicle by unscrewing the single wheel
nut. In our design we aimed at implementing an
encoder system that would be able to withstand
variations in alignment and sensor-to-wheel
spacing. Because the encoder is open to the
environment, we chose to avoid optical slit or
reflective surface sensing, or any system that dirt
or grime could interfere with. For these reasons
we chose a solution based on magnetic sensing.
We designed custom encoder disks to mesh
with the structure of the Segway wheel hubs. We
then laser-cut these disks out of durable acetal
plastic (see Figure 2). Each encoder disk has 100
holes around its rim, into which we press-fit
cylindrical neodymium magnets. The magnets
are axially magnetized, and placed with their
magnetic fields in alternating orientations. This
way a sensor at the rim of the disk will see
alternating magnetic poles as the disk rotates.
EXPERIMENTAL RESULTS
This section presents experimental results obtained with a Segway HT that was retrofitted with
the odometry system described in Section 2. We performed 10 indoor runs, each lasting just over
15 minutes and averaging 1,880 meters (1.16 miles) in length. In each run, the Segway started
and stopped at the exact same location. Figure 6 and Figure 7 show the results of two
representative runs. In order to create these figures, we plotted each estimated trajectory over the
appropriate floor plan. For each run, we show the trajectory resulting from odometry only, and
from odometry plus HEO corrections. As is quite apparent from these qualitative results, with
HEO there are virtually no heading errors.
In order to obtain quantitative experimental results, it is necesary to compare the measured
position and heading against ground truth. In outdoor experiments, ground truth can usually be
obtained conveniently from GPS. Indoors, without GPS, obtaining ground truth is a bigger
problem. For our indoor experiments, we compiled ground truth data for X-Y position and
heading from architectural floor plans. Using this rather tedious method, ground truth for X-Y
coordinates can only be estimated by assuming the Segway drove along the center of corrdidors.
In reality, and especially in wide corridors, the rider frequently switched “traffic lanes” to avoid
pedestrians. Furthermore, we did not estimate ground truth position and heading during turns,
since it is virtually impossible to guess at the Segway’s true trajectory and momentary heading
during turns. For the same reason, we did not attempt to compile ground truth heading data
during quick lane changes or obstacle avoidance maneuvers.
Arbitrary Motion
In order to get a sense of how sensitive the HEO algorithm is to deviations from the heuristic
assumption of travel along dominant directions, we performed two additional runs. In run
“Arbi-1” the Segway rider performed arbitrary maneuvers such as driving around in circles for
three 30-second intervals during a 15-minute ride (i.e., for about 10% of the total trip time), as
shown in Figure 10. Besides these three intervals, all motion was along dominant directions, as
in the 10 original runs. All odometry and HEO parameters were kept exactly the same as in the
10 original runs, that is, as shown in Table II and Table III. The results of Arbi-1 with HEO
correction were an average heading error of 0.65° and an average position error of 1.7 m. Both
results are comparable to those of the 10 original runs.