At the forefront of Artificial Intelligence

Home » Articles » Neural Networks » Applications/Code

Mouse Gestures Recognition

By Konstantin Boukreev

Title:       Mouse gestures recognition
Author:      Konstantin Boukreev
Email:       konstantin@mail.primorye.ru
Environment: VC++ 6.0. SP5, Win2k SP2, MS Platform SDK April 2001
Description: Feedforward multilayer neural network and mouse gesture recognition

Sample Image

Introduction

Recently I installed Opera 5 and was impressed on a Gesture UI. IMHO, the neural network most suitable for this purpose. As I a little know neural network I tried to implement such feature themselves.

Neural Network

What is Neural Network ? Hm it's not easy to say. A rephrased definition Zurada, J.M.:

"Neural network software is a software which can acquire, store, and utilize experiential knowledge."

I think I can point any person concerned to theory directly to several neural network sites. Here is small list of web resources about Neural networks:

FAQ of comp.ai.neural-nets
Generation5
"Neural Computing: Theory and Practice" 1989 Philip D. Wasserman (in russian only)

Implementation

Let's return to mouse gestures. After some research I have chosen a multilayer perceptron and standard back-propagation algorithm for training. The main problem was in the representation of an input data for neural network. The best result I found was in the transformation of a mouse path into a vector of cosines and sines.

For example:

path   {170:82 172:83 175:85 177:86 ...} 
transformed into 
vector {0.45 0.55 0.45 0.71 0.89 0.83 0.89 0.71 ...}

Recognition algorithm.

record a mouse path
smooth a path to a base points
transform points to angles' vector
compute sines and cosines
pass values (cosines and sines) to network's inputs
apply softmax function on an output network vector
find and verify a winner

Neural network architecture.

input layers : 32 sinapses
hidded layer : 32 neurons
output layer : 29 axons (one for each gesture)
fully connected layers
transfer function : log-sigmoid
incremental training algorithm, standard back-propagation method
momentum, variable learning rate (slowly reduced)
input noise

Application

Training

Sample Image

Before testing the recognition ability you must train the network (or you can load an file image of trained net). You can customize the parameters of the training process, namely: maximum number of cycles, a momentum value, a learning rate, a minimum value of mean square error (in other words "target error"). The training process will stop after achieving either of the conditions: maximum number of cycles or target error. During the training process you can keep an eye on a error's graph, a current gesture (with noise) and 2D network presentation.

Testing

As soon as you have a trained net, you can test it. Select the patterns (or test all of them), a speed value and a noise level. Besides, you can familiarize oneself with ideal presentation of gestures via setting minimal noise and minimal speed.

Recognition

For recognition of mouse gestures you must press right mouse button during moving a mouse. For example for recognition "left" gesture, press right mouse button and move a mouse to the left. If a neural network can recognize the gesture, then you will see the name, probability and ideal presentation of winner. Because of freeware nature of GestureApp the mouse path must have at least 16 points :(. Sorry I didn't implemented a "stretch a path" feature so far.

Note: the direction is very important.

The network is trained to recognize the gestures but not 2D images. Hence, you can draw the "circle" gesture a thousand different ways, but the only valid way is: press mouse button and move a mouse to the right and down and so on. Once more: it's gesture, not 2D image.