dc.description.abstract | With the increasing usage of new technologies in common daily activities, the demand
of efficient human-computer interaction (HCI) systems increases. Hand pose recognition
systems have been widely explored for such task due to its intuitive operation for non
experienced users. However, vision-based hand pose recognition is a extremely challenging
problem due to the dynamics of the hand, which poses a large amount of degrees of
freedom that makes it difficult to estimate and carries out additional problems such as
self occlusion. With the development of reliable and consumer affordable vision systems
such as the Microsoft Kinect©, depth imaging has become a useful tool on body parts
recognition and thus, for hand recognition.
This thesis proposes a static hand pose classification system based on depth images only
and considering a top view perspective. No additional constrains to the hand position
on the scene are imposed, which allows background objects to be closer to the camera
than the hand itself. A synthetically generated data set of four hand postures (open,
pointing, fist and pinch) is used. The proposed design is divided in two processing stages:
hand segmentation and hand pose classification. The hand segmentation stage uses a
random decision forest (RDF) for per-pixel classification of the depth images, segmenting
the hand in arm, palm and fingers regions. Hand pose classification is then performed
using a defined set of visual features from the labeled blobs. Seven visual features are
evaluated in terms of classification accuracy. Two types of classifiers are trained for
the pose estimation: random decision forests and support vector machine (SVM) for
evaluation purposes. The system proposed provides a 91% of classification accuracy for
the defined hand poses on the generated data. | es |