|
[DPRG] progress on Hough transform and small robot
Subject: [DPRG] progress on Hough transform and small robot
From: Chris Jang
cjang at ix.netcom.com
Date: Sun Jan 21 01:24:52 CST 2007
Hello,
The basic Hough transform code works but is even less reliable than the
object recognition of soda and beer cans. It needs work.
http://golem5.org/embedcv/images/hough.jpg
The bright points in the Hough transform image correspond to the detected
red lines in the (histogram equalized) edge image.
One thing I've noticed in real world images is that horizontal edges tend
to be much stronger than vertical or diagonal orientations. My theory is
that this is an optical effect. Horizontal edges tend to remain within the
camera's depth of field under even lighting.
I'm also starting to think that thresholding the Hough transform of a
Sobel edge image is too difficult. The algorithm is very sensitive to
noise and does not reject it very well. Better edge detection than vanilla
linear convolution is required.
Another thing I'm going to do is test EmbedCV on a 1/12 scale RC car. Here
is a picture of it.
http://golem5.org/embedcv/images/littlebro.jpg
It cost $15 (unsold Christmas product) and has full suspension (not much
travel but all four wheels are sprung) and a rear differential. There is
no PWM and steering is full deflection bang-bang. The wheels are rubber.
It is just wide enough that a PC 104 SBC can fit inside the body. I'm
thinking of the Technologic TS-7260 ARM9 board. The battery pack voltage
is stable enough that I think everything can run off the same pack without
isolation between the motor drive and digital electronics. A notebook
webcam attached to the roof provides a camera.
This toy is small enough to fit in a shoebox and so can run on the DPRG
warehouse course area. It's big brother is about 10x heavier and barely
fits in the trunk of my car. I have not been able to make much progress
with it as the logistics of testing are involved. I'm hoping that less is
more and this smaller robot allows for accelerated development.
One thing that amazes me is how much faster buying off-the-shelf is. It
took months to build a frame, drivetrain, motor control electronics and
electronics enclosure for the big robot. If it proves easy to hack the
motor control of this toy and all electronics can just be stuffed inside
it, then there's easily an order of magnitude improvement in development
time. This isn't quite true - cost has been shifted from hardware to
software. I've been working on EmbedCV for several months and it is still
far from working on a robot. If this works, it may be a better time/money
tradeoff for many people who wish to build robots - COTS vehicle platforms
with 32 bit embedded computers and more advanced signal processing for
sensing the environment.
My theory is that PWM speed control added to the rear drive is enough that
bang-bang steering is ok. I think that driving in smooth curves is
somewhat difficult. Most robots drive in straight lines. They only turn
when pointed in the wrong direction. That's how I'm viewing this robot.
When turning, it can just slow down so the vision and control system can
keep up with the angular rate of turn. Then once pointed in the right
direction, it can increase speed.
Here's a better photograph of the electronics board inside the toy.
http://golem5.org/embedcv/images/img3868.jpg
If anyone has some experience regarding hacking this kind of thing, I
would appreciate your advice. I plan on using the DIO lines (appropriately
level shifted and buffered) from the ARM9 SBC to tap into the existing RC
circuits and control the car. I'd like to leave the radio receiver
electronics in place too. Right now, I am thinking absolutely minimal
modification.
Some things I recognize. Those three big diodes must be for motor flyback
voltage. There's a coil and IC for the radio. Does anyone know what that
company insignia is? Where can I find a datasheet? There are a whole bunch
of bipolar transistors for motor control. I didn't expect this many. I'll
reverse-engineer and diagram the circuit soon.
Chris
More information about the DPRG mailing list
|
