Hand gesture recognition for human-robot interaction

Abstract

This project proposes an image segmentation method that improves the recognition rate of vision-based hand gesture recognition system on low-resolution images and occluded hand gestures. The solution is based on the idea that random walker-based segmentation could provide high quality segmentation despite weak object boundaries. The approach has several notable merits, namely high segmentation accuracy, fast editing and computation. A comprehensive verification using Matlab is carried out to determine the effectiveness of random walker method in segmenting occluded hand gesture images. The segmented images are then classified by artificial neural network and its performance is evaluated in terms of recognition rate and time. The result confirms that the proposed method is performs better than color-based segmentation, that is 5% higher recognition rate for the same dataset. The method proposed in this project can be integrated in vision-based recognition systems to widen the vocabulary of hand gestures recognition systems, recognizing both gestures with finger gaps as well as occluded gestures.