Comments by Matty22 (Top 5 by date)

I believe you'll have to
Train Kohonen unsupervised on a bunch of videos some with obstacle and some with no obstacle. With osbstacle in different positions.

This will create a clustered map with some neurons correlating to no obstacle and some correlating to obstacle

You then have to look up test image in the map and decide based on its position on the map

Because SOM is unsupervised learning you've got no control as to how it clusters things.

However that doesn't preclude generating a 'map' with the SOM and using some other technique to look up a test image from the camera in the map to decide if there is an obstacle or not


Have a look at this paper
http://metronu.ulb.ac.be/imacs/papers/T3-I-112-1018.pdf

Have a look at Fig 8 in that document.

In that paper seems like they use a SOM to create a map clustered into regions like 'object in middle, object in far left, object in top right, no object' etc. Then in order to perform the detection they look up the current image in the map produced by the SOM.

But keep in mind the SOM itself isn't being run/executed to determine if there's an obstacle here. The SOM only produced a training map that you have to interpret in code. This is because SOM training is unsupervised, you have no control over what output neuron will mean 'no obstacle'

This is different from say a perceptron or hopfield neural network where you use supervised training to tell the network to turn a neuron to 1.0 if there is an object and -1.0 if there is none. With the SOM you have to write code to use the map as a guide. With a supervised trained neural network, the neural network can give an answer to questions like 'is there an object in the way?' without having to look up stuff in a map like with a SOM. The SOM always has many many outputs that you have to interpret somehow....with a more traditional neural network you can just have 1 output, 'is there an object in the way?'

To my knowledge there's no method you could use to make a SOM do obstacle detection.
You can use a SOM to create a training set that have IMPORTANT vectors (hence using the clustering component of the SOM) Then use some OTHER method to determine if there is a match

But by itself. SOMs have no more to do with detecting objects than Boyle's gas law does.

The strongest object recognizable systems don't even use neural network or AI at all.
They use eigenfaces/eigenfeatuers and principle component analysis

http://en.wikipedia.org/wiki/Eigenface

Other options are Hopfield networks, edge detection, etc.

But yeah..you can probably forget about using a SOM for anything like this...not really what they do. (Principle component analysis in eigenfeatures is kind of related to what a SOM does (vector reduction) But it's only a small part of the problem of 'object recognition' You COULD use a SOM as a component in a object recognition system..but it would only be a very small part that's used to help prepare the training set for another system that does the actual recognition.

Kohonen SOMs are, like the name suggests. A map that self organises, eg, clusters, reduces vectors. You CAN use them to do processing of video streams...if there's something in the stream you want to cluster/reduce vectors of. If your trying to use a SOM to recognize obstacles or find edges or see a face then you're using entirely wrong algorithm. A SOM can't do this by itself