|
[DPRG] Placement and interpretation of ultrasonic
Subject: [DPRG] Placement and interpretation of ultrasonic
From: dpa
dpa at io.isem.smu.edu
Date: Wed Jun 6 15:50:05 CDT 2007
Howdy,
chris writes:
>>Only problem with this method is that it can be difficult to tell the
>>difference between the reflections from each pulse. Pulses return via
>>multiple paths from everything in front of the sensor, not just the
>>nearest object.
>>
>>Some of the more complex robots (and ground radar) exploit the multiple
>>return pulses to build their map faster, and I've read a paper
>>( http://www-personal.umich.edu/~johannb/Papers/paper32.pdf ) that
>>'scatter-fires' the sonar units in a pseudo-random fashion, then applies a
>>best-fit algorithm to the results in order to differentiate between
>>reflections from the associated tx and other txs with an overlapping field
>>of view or multiple reflection paths around the robot.
The sonar array for jBot was orginally designed for 6 sonar arranged
at 15 degree spacing for a total frontal coverage of 90 degrees.
However, the current configuration uses only 4 sonar for a total
frontal coverage of 60 degrees.
This allows the 4 sonar to be fired in sequence with a total wait
time of 1/16 second. That allows the maximum range of the sonar,
which is 32 feet, or 64 feet total travel time from the sonar to
the object and back. So multiple echos from any sonar ping have
all decayed before the next sonar is activated. That gives 4
readings per second for each sonar.
Going to an array of more than 4 sonar either requires fewer
measurements per sonar per second, which greatly degrades navigation
at 5 mph (jBot top speed), or using something like Borenstein's
algorithm, which requires much more complex signal processing. So
I opted for 4 sonar, and it seems to be sufficient.
I have run into problems with very oblique angles of highly
reflective surfaces like glass, which would probably go away
with the full 90 degree coverage. My current scheme is to solve
those problems with a different sensor technology: two close-in
IR rangers angled out +- 45 degrees from center. But that's
just conjecture at this point.
>I had been envious of pulsed time of flight rangefinding systems like
>LIDAR and SONAR as they appear so much easier than vision. I had thought
>that clear distance maps would just come out as by magic. But in two
>cases, LIDAR on Stanford's Grand Challenge robot and SONAR from this
>paper, some form of stochastic inference was necessary. The raw sensor
>data did not make sense. Also, in both cases, time-domain filters were
>used (e.g. Kalman filter) to update a belief model.
jBot uses its sonar for navigation and obstacle avoidance, as well as
perimeter following. It does not, however, attempt to generate an
internal map of the environment based on those sonar readings. My
experience is that approach in general is not a very robust method
of robot navigation, and difficult to do in realtime.
best regards,
dpa
More information about the DPRG mailing list
|
