Vision-based hand shape identification for sign language recognition
Abstract
This thesis introduces an approach to obtain image-based hand features to accurately describe hand shapes commonly found in the American Sign Language. A hand recognition system capable of identifying 31 hand shapes from the American Sign Language was developed to identify hand shapes in a given input image or video sequence. An appearance-based approach with a single camera is used to recognize the hand shape. A region-based shape descriptor, the generic Fourier descriptor, invariant of translation, scale, and orientation, has been implemented to describe the shape of the hand. A wrist detection algorithm has been developed to remove the forearm from the hand region before the features are extracted. The recognition of the hand shapes is performed with a multi-class Support Vector Machine. Testing provided a recognition rate of approximately 84% based on widely varying testing set of approximately 1,500 images and training set of about 2,400 images. With a larger training set of approximately 2,700 images and a testing set of approximately 1,200 images, a recognition rate increased to about 88%.
