29-03-2012, 04:26 PM
Orientation Histograms for Hand Gesture Recognition
10.1.1.6.618.pdf (Size: 302.76 KB / Downloads: 69)
Introduction
Computer recognition of hand gestures may provide
a more natural human-computer interface, allowing
people to point, or rotate a CAD model by rotat-
ing their hands. Interactive computer games would
be enhanced if the computer could understand play-
ers' hand gestures. Gesture recognition may even be
useful to control household appliances.
We distinguish two categories of gestures: static
and dynamic. A static gesture is a particular hand
conguration and pose, represented by a single im-
age. A dynamic gesture is a moving gesture, rep-
resented by a sequence of images. We focus on the
recognition of static gestures, although our method
generalizes in a natural way to dynamic gestures.
Our Approach
We seek a simple and fast algorithm, which works in
real-time on a workstation. We want the recognition
to be relatively robust to changes in lighting.
A high level approach might employ models of the
hand, ngers, joints, and perhaps t such a model
to the visual data. This approach oers power and
robustness, but at the expense of speed.
A low-level approach, such as was taken by [6],
would process data at a level not much higher than
that of pixel intensities. This approach would not
have the power to make inferences about occluded
data. However, it could be simple and fast. We chose
this approach.
Operation
Figure 4 illustrates operation. There is rst a train-
ing phase. The user rst indicates the hand posi-
tions for the desired vocabulary of gestures, such as
the commands for \up", \down", \right", \left" and
\stop" in this example, (a). We show only 3 com-
mands in the gure, but typically more are used. The
user may show several gesture examples correspond-
ing to a single command. The computer stores the
orientation histograms corresponding to each image,
Problem images
From our experience watching many people use the
system, we have observed several conditions where
the user is not satised with the gesture classication.
These are illustrated in Fig. 6.
(a) and (b) show two images which many users feel
should represent the same gesture. However, their
orientation histograms are very dierent, ©. In the
present system, this problem can only be addressed
by providing multiple training images for the same
gesture.
Conclusions
We have applied a simple pattern recognition tech-
nique to the problem of hand gesture recognition.
For static hand gestures, we use the histogram of lo-
cal orientations as a feature vector for recognition.
This method has a training phase and a run phase.
In the training phase, the user shows 5 to 15 example
hand gesture commands. The computer stores one or
more feature vectors, blurred orientation histograms,
for each command. In the run phase, the computer
compares the feature vector for the present image
with those in the training set, and picks the category
of the nearest vector, or interpolates between vectors.