in
Welcome to Lionhead Community Sign in to Windows Live ID | Help

ANN for gesture recognition

Last post 05-12-2009, 14:35 by Kindjie. 4 replies.
Sort Posts: Previous Next
  •  06-15-2006, 15:37 2005106

    ANN for gesture recognition

    hi. i've been reading up all day and am gettign ready to implement a neural net for fairly trivial gesture recognition and was wondering if anyone had any tips. i noticed the ANN thread but it's too long to read through and going into much more detail than required here. we're planning to implement a 3 level backward propogated nn for the task. what do you think? also. if you have any resources you can suggest we should be able to track down books and the like. cheers imp
  •  06-17-2006, 18:16 2012595 in reply to 2005106

    Re: ANN for gesture recognition

    Well, don't fix the geometry of your network. With a proper design, changing the number of layers and the number of neurons per layer should be a matter of changing a few lines and recompiling. Also, I think the representation of the gestures is going to be the most important thing to get right. Think carefully about that.
  •  06-19-2006, 10:20 2017259 in reply to 2005106

    Re: ANN for gesture recognition

     Quoting: Lourens Veen
    Well, don't fix the geometry of your network. With a proper design, changing the number of layers and the number of neurons per layer should be a matter of changing a few lines and recompiling.
    Totally Ditto with that - you might end up wanting to fiddle about with the structure of your NN a lot and it would be nice if you could do it easily. If I was doing this in the way you suggest, I'd have a totally generic way of representing the NN and then create a Genetic Algorithm to work out the structure for me. -- If I was specifically doing Gesture recognition however, I'd either be using a Kohonen network ( Unlikely ) or a simple single layer of perceptrons for each gesture and maybe another layer for noise removal.
    I have no idea what I am talking about
  •  06-22-2006, 1:03 2026337 in reply to 2005106

    Re: ANN for gesture recognition

    Well, I'm implementing some functionality in my neural net at the moment to let it recognize gestures, since I've already finished the network I suppose I could have some tips at hand. Representation of the gestures is pretty inituitive. Throw in an array of x & y delta's ( x(n-1) - x(n), not plain coordinates!), tell the network wich pattern you're throwing at him, and the network will eventually recognize the patterns. While I've not tested this with true mouse-gestures, this seems to work with pre-generated patterns (with some \ alot noise added to it) so I figure it'll do just as fine with the mouse. What Laurens said about keeping things flexible is very true. You don't want to alter your entire program just because you've figured out that you need 30 more neurons, or more layers. Some understanding of basic datastructures (linked lists etc) comes in handy here. First get your datastructures right, after that, program the output functions for the neurons, than start with the backpropagating algorythm wich is fairly complicated if you're just starting (I myself only have vague understanding of how it works - exactly - ) Usefull read: http://www.generation5.org/content/2002/bp.asp If you have any specific questions about how I implemented the NN, just ask!
    Hapiness isn't hapiness without a violin playing goat
  •  05-12-2009, 14:35 3344522 in reply to 2005106

    Re: ANN for gesture recognition

    Hi,

    I built one to recognize Wii gestures a while back. It worked quite well despite being rushed, although I wouldn't use my implementation for realtime games. Here's a link to my paper: http://kindjie.googlepages.com/

    You can skip the Wii specific stuff. The actual code for all of it was only a few dozen lines, so my advice is to keep it simple. Smily [:)]

    G o o d [Good] luck!
View as RSS news feed in XML