Just a thought, maybe you should start with a small B/W bitmap image file and see if you can process what it contains line by line.
zoomkat:
Just a thought, maybe you should start with a small B/W bitmap image file and see if you can process what it contains line by line.
Yes before buying all the hardware, I am going to start experimenting with processing and various images.
Thanks for the help. I guess I've figured out what exactly has to be done. I will ask again if I get stuck somewhere specific.
HappyTron:
Okay, nevermind. In India the Kinect for PC is costing around $400.
Wow, that is expensive.
I bought it in Indonesia, only cost around US$250..
Hmm it might be a little bit impossible to get the image data from kinect wirelessly since, it require very high data speed, remember if you wanna operate Kinect for Windows in PC, you need to install a lot of things, Simple OpenNI, OpenNI, KinectSDK, etc. It is my opinion, correct me if I am mistaken.
It is easy to follow light, if you turn all the rest of the lights in the room off.
How reliable does this thing have to be ? Work outdoors, or indoors only ? Can you guarantee that there
will be no other "red" things in the room for it to home in on ?
Arduino is not really capable of real "machine vision" capability.
One idea, might to be make your led blink at some specific and unusual frequency ( not necessarily visible),
and set up an antenna on the robot which is going to notice that specific frequency and home in on it.
Points at grumpy Ha Haaaa!
Beat ya! 
Before the TSR80, with a home made camera, a home designed/built Z80 'pc' and a neural net; I was able to identify if the "camera" was looking left, right or centre of a scene.
The camera used a loop of super 8 film (hand cranked, never motorized it) as a pixel mask to give a 16 and then a 64 pixel (8x8) device (just filmed a white card on a black background to create a moving window) and a single photo transistor 
No forums (or web) back then The point being, don't let them tell you it's impossible, keep plodding and use your imagination. There are many ways to the solution, but the fastest way there is to understand the problem. 
I was talking to a guy a couple of years ago who was into neural nets, and he said he uses high end video cards, and hacks the multiple cores/processors as a very fast engine to "parallel process" his nets.
Now matter how fast the processor, we will always overload it with crud and make out it's slow! Like cupboard space, you always have 10% more rubbish than space!
I know now!!!! Used an IR led flashing at 38 khz and put an IR receiver on the Arduino robot!
HappyTron:
It will be image processing cause the light which the bot has to follow, I will be holding it in my hand waving around in a 3-D space.
So you have a light, which the 'bot has to follow. Could you explain how you came to the conclusion that this is not a light-following problem?