5.2 Getting a clean signal

It was found that while the information coming from the fingers was reasonably clean because of the automatic calibration that occurred between each sign, but positional information (in terms of x, y and z) was often found to ``glitch''; that is, to change to values that were not just randomly noisy, but physically impossible. There was of course much noise on the position (as would be expected with 8-bit resolution), but most of it was no more than 1 to 2 centimetres of displacement. Occasionally, however, the data returned would indicate that the hand had moved 40 or so centimetres in 1/25th of a second, which would hardly ever occur in the course of normal signing. There are several causes for this:

• Random ultrasonic noise. Things like monitors and fluorescent light bulbs generate some ultrasonic frequency noises. Also, and this is probably the more problematic, reflections of the signals emitted by the glove itself off rigid, non sound-absorbing surfaces will interfere with the positioning. This is remedied by the fact that most commonly the original signal will be the strongest.

  Steps were taken to reduce the effect of noise. For example, the environment was arranged so that the walls were at a angle so that reflections were not likely to interfere. The walls were also covered with a thin material which reduces the intensity of the reflected signals. Tests with the fluorescent lights on and off showed that they had little effect on performance.

• Ultrasonic sensors work best in ``line of sight''. This means that if an object blocks the path between the glove and the L-bar, the glove becomes more susceptible to glitches, since it is possible in this case that a reflection is stronger than the original signal. This can happen quite frequently, most commonly when the hand itself gets in the way. For example, consider waving hello. The hand blocks the space between the transmitter and the receiver. This effect is even more dramatic when the transmitter and the receiver are pointing in opposite directions, in which case the data coming in is almost random. This also happens frequently -- consider the sign for I where you point towards yourself. There is very little you can do about this.

To remove these glitches a heuristic method was adopted. This is best expressed in a mathematical notation.

Each file analysed consists of a sequence of frames. In each frame there is information about x, y and z position, wrist rotation and finger bend.

Let be the x-position, y-position and z-position stored in the ith frame of the sample.

Then we can define the change in direction from the last frame as:

The vector is thus a pointer in the direction we are going in. Thus if we take the length of this vector and call it , i.e.:

then this will give us a measure of speed. In effect this is the first (discrete) derivative with respect to time.

Similarly we define the second derivatives:

Here again, the vector is a vector that represent the change in direction that has occurred. is similarly the length or norm of that vector.

Thus an intuitive way to filter the glitches is:

1. See what the practical limits of movement are. This was done by creating special signs -- such as fast movement in a straight line, making the hand move in small tight circles in a tight manner, and rapid to and fro movement. These files were ``hand de-glitched'' using the gloveplay program on the data files. They were then put through the analysis discussed above to find what the practical limits were.
2. When filtering a sign, check the and . If these exceed the practical limits, then average the previous point with the next point and set the value of this point to that average.

This filter is very empirical and the careful eye will observe some problems with it. For example, what happens if you get two glitch readings in a row? Clearly the above algorithm assumes that the next value after a glitch is a sensible one. It was found empirically that two glitches in a row rarely occurred. Furthermore, even if it did occur, the glitches would still be smoothed over, but to a lesser extent.

In fact, this can be thought of as nothing more than a specialised threshold filter.

Other filters were considered, such as Kalman filters, moving average filters and so on. However, these were not implemented for a number of reasons:

• Their exact meaning in a three-dimensional space is not always clear. For example, is the filter applied to the three dimensions separately, or is there some way to map these methods more effectively to deal with all three variables at one time?

• Most of these filters are designed to smooth out random noise (which is usually assumed to be normally distributed). While there was certainly random noise to a certain extent, most of the feature extraction techniques would prove robust against such random noise. However, glitches were another story altogether, they are not distributed in the same way that the smoothing filters expect. The features used were found to be more sensitive to glitches than noise. For example, a glitch could easily throw off a bounding box, since the box would grow to accommodate the glitch point. Similarly if some distance or energy measurements were to be used, a glitch (if it was to happen) would represent a giant burst of energy relative to the rest of the motion.

• They were found to be computationally complex, while the filter proposed above is computationally simple. Furthermore, as shall be seen, a number of the synthesised features use the values calculated above anyway. If the filter was applied at the same time as the features were being extracted there would be no significant penalty.

This was implemented, as was all the feature selections in simple perl scripts.

No major sort of fitering of outlier data samples was taken, although some of the data that was incorrectly collected (such as one example where the user had taken a drink and forgotten to press ``B'' to stop data capture, resulting in a file with over 600 frames in it). This occurred four time in the course of sampling.

We will now go on to discuss and assess the features that we extract.