Person Independent Recognition Of Dynamic Hand Gestures Using Vision Based Systems

W hen a sign language recognition system will be used for teaching purposes, the students cannot train the system. Therefore person independent recognition has to be performed by using a pre-assembled gesture data set. Although many systems have been developed that train and recognize the same person, few research is done on person independent gesture recognition. In this paper we present three different visual gesture recognition systems, and compare their performances. The first system uses a naïve Bayes classifier in which the continuous valued feature vectors are modeled using Gaussian distributions. The second system is similar but uses a quantized version of these feature vectors. The last system is based on Hidden M arkov M odels. W e discuss how the hands are detected in the images. The features that are used for recognition are the mass, position, rotation, perimeter, length to width ratio, velocity and acceleration of the hands in a two-dimensional image. These features are normalized to the size of the head. In the experiments the best features are selected and the recognition performance of the systems is optimized using a forward gready search through all features. Besides several parameter optimizations have been done for the three classifiers. W e obtained a maximum person independent recognition rate of 90.7% for the gaussian-based naïve Bayes classifier, 94.4% for the discrete naïve Bayes classifier and 80.6% for the HM M approach. The experiments were performed using a dataset containing twelve different gestures, performed by five different persons.