Hand Gesture Recognition Using Artificial Neural Networks

Hand gesture has been part of human communication, where, young children usually communicate by using gesture before they can talk. Adults may have to also gesture if they need to or they are indeed mute or deaf. Thus the idea of teaching a machine to also learn gestures is very appealing due to its unique mode of communications. A reliable hand gesture recognition system will make the remote control become obsolete. However, many of the new techniques proposed are complicated to be implemented in real time, especially as a human machine interface. This thesis focuses on recognizing hand gesture in static posture. Since static hand postures not only can express some concepts, but also can act as special transition states in temporal gestures recognition, thus estimating static hand postures is in fact a big topics in gesture recognition. A database consists of 200 gesture images have been built, where five volunteers had help in the making of the database. The images were captured in a controlled environment and the postures are free from occlusion where the background is uncluttered and the hand is assumed to have been localized. A system was then built to recognize the hand gesture. The captured image will be first preprocessed in order to binarize the palm region, where Sobel edge detection technique has been employed, with later followed by morphological operation. A new feature extraction technique has been developed, based on horizontal and vertical states transition count, and the ratio of hand area with respect to the whole area of image. These set of features have been proven to have high intra class dissimilarity attributes. In order to have a system that can be easily trained, artificial neural networks has been chosen in the classification stage. A multilayer perceptron with back-propagation algorithm has been developed, thus the system is actually in-built to be used as a human machine interface. The gesture recognition system has been built and tested in Matlab, where simulations have shown promising results. The performance of recognition rate in this research is 95% which shows a major improvement in comparison to the available methods