Principle

We can similarly apply histograms to the wrist roll and finger position. This is still an intuitively useful variable, since, for example, it tells us how much time the hand was at a particular roll level. Clearly signs such as give with the palms up, will have rotational histograms that significantly differ from those that have the palms down, such as take. Similarly, I, with a pointing gesture to oneself, will have a different finger histogram (hopefully) to mine which is a similar gesture to I except it is with a clenched fist.

They differ from the x, y and z position histograms in a number of ways, however:

• The finger and rotation information is not as high resolution as the x, y and z information (for example, x y and z position each have 256 possible values; rotation has 12 and fingers 4 each). This means that histograms can be built on each of the raw data values -- rather than having to divide the range of values up.

• The idea of normalising by maximum and minimum values of the variables does not apply. For example, the rotation value indicates a particular angle between 0 and 360 (at 30 intervals). We want this in absolute terms, not relative terms. Similarly, because a value of 0 means a finger is fully extended and 3 means fully flexed, we want these pieces of information.

There are some special precautions we need to take, as well as some redundancy we can make use of.

  
Figure 5.4: The possible rotation values coming from the glove and their encoding and classification into histograms.

In the rotation histogram, there are 12 possible values, as shown in figure 5.4. Although we could directly build histograms on this data, this would make the system extremely sensitive to noise, since practical observation of the glove indicates that the ``roll'' value is particularly sensitive to noise. Furthermore, it is unlikely that a sign will intentionally be made at an angle not parallel to one of the axes. Few, if any, signs, for example, require the hand to be inclined along a 45 angle.

Thus what we can do is segment the region into areas as shown in figure 5.4 -- into 4 regions, centred on the positive and negative x and y axes. This gives us a much more noise-resistant classification. For values of 11, 0 and 1 we classify this as ``palm down''. Similarly values between 5 and 7 would indicate ``palm up''. Similarly for ``palm left'' and ``palm right''. Histograms are then built on the proportion of time spent with palm up, palm down, palm left and palm right.

For the fingers, we can make a similar suggestion (in fact Sturman [Stu92] did) -- that 90 per cent of the time, the fingers are either fully opened or closed. While this method works in cases with high-accuracy gloves, in this case, there is no margin for error with only two bits of information available. Also, since there are only 4 possible values, it is practical to have a histogram for each finger position.