Summary (long): pattern recognition comparisons

[Dean.Pomerleau@F.GP.CS.CMU.EDU]

Leonard Uhr writes

> Neural nets using backprop have only handled VERY SIMPLE images, usually in
> 8-by-8 arrays.

and later

> What experimental evidence is there that NN recognize images as complex as those
> handled by computer vision and pattern recognition approaches?


For the past two years I've been using backpropagation networks with
32x30 and 45x48 pixel retinas and up to ~20,000 connections to autonomously 
drive a Chevy van.  This system, called ALVINN (Autonomous Land Vehicle
In a Neural Network), uses a video camera or 2D scanning laser rangefinder
as input, and outputs the direction in which the vehicle should steer.  The
network learns by watching a person drive for about a 1/4 mile.  After
about 5 MINUTES OF TRAINING, the network is able to take over and continue 
driving on its own.

Because it is able to learn what image features are important for 
particular driving situations, ALVINN has been successfully trained to 
drive in a wider variety of situations than any other single autonomous
navigation system, all of which use the traditional vision processing 
techniques Leonard Uhr refers to.  The situations ALVINN networks have
been trained to handle include single lane dirt roads, single lane paved 
bike paths, two lane suburban neighborhood streets, lined two lane highways, 
and, using the laser range finder as input, parking lot driving.
Because of its ability to effectively integrate multiple image features 
into a single steering command, ALVINN has proven more robust than other
autonomous navigation systems which rely on finding one or a small
number of features (like a yellow road center line) in the image.

Because of the simplicity of the system, it is able to process up
to 29 images per second (both training and testing are done using
two Sun-4 Sparcstations).  ALVINN is currently limited in the speed
it can drive by the test vehicle, which has a top speed of 20 MPH.

Autonomous navigation was one domain in which traditional vision
researchers were initially skeptical that artificial neural networks 
would work at all, to say nothing of work as well or better than other
systems in a wider variety of situations.

--Dean

Pomerleau, D.A. (1989) Neural network based autonomous navigation. In 
Vision and Navigation: The CMU Navlab. Charles Thorpe, (Ed.) Kluwer
Academic Publishers.

Pomerleau, D.A. (1989) ALVINN: An Autonomous Land Vehicle In a Neural
Network, Advances in Neural Information Processing Systems, Vol. 1, D.S.
Touretzky (ed.), Morgan Kaufmann.