Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Videobasierte Objekt-Verfolgung mit "Hidden-Markov-Modellen"

: Reinhardt, J.
: Malerczyk, C.

Friedberg, 2002
Gießen-Friedberg, FH, Dipl.-Arb., 2002
Fraunhofer IGD ()
tracking; Hidden Markov Models; interaction; virtual reality

Easy and intuitive interaction with virtual worlds is still an important task in computer graphics. Creating a completely device less interaction is a step leading us to a very flexible method to control and manipulate wireless with virtual worlds, even for first time users. In this diploma thesis a method based on Hidden Markov Models to track a pointing gesture within sequences of grey-scaled images is presented. A gesture is represented by a set of landmark points located on its contour. Possible deformations of this contour (the modes of variation) are statistically described by a Point Distribution Model (PDM). Due to the fact that the PDM is operating on high dimensional spaces, a prediction on the PDM's data is difficult and time consuming and therefore the bottleneck in a real time application.
With the Viterbi algorithm we have a powerful method at hand to use Hidden Markov Models for gesture tracking. The Viterbi algorithm leads from a given sequence of observations to the most probable sequence of underlying states. During a learning phase a Hidden Markov Model is created by calculating conditional probabilities between states and observations, which are interdependent. In the case of the gesture tracking we identify the 2d position of the gesture (i.e. the tip of the index finger or its centre of gravity) in image space as an easy to handle and simple to predict observation. This position is directly depending on the appearance of a gesture (the deformation of the contour) due to a given camera set-up. Predicting the easy to handle 2d translation by applying a non-linear regression leads an extended vector of observations to a corresponding vector of states. The last element of the vector is a predicted contour, which is used as a rough approximation of the gesture in the next image.