Classification of visual hand movements using multiresolution wavelet images

This paper presents a novel technique for classifying human hand gestures based on stationary wavelet transform (SWT). It uses view-based approach for representation of hand actions, and artificial neural networks (ANN) for classification. This approach uses a cumulative image-difference technique where the time between the sequences of images is implicitly captured in the representation of action. This results in the construction of motion history images (MHI). These MHI's are decomposed into 4 sub images using SWT, approximate and detailed images. The approximate image is fed as the global image descriptors to the ANN for classification. The recognition criterion is established using backpropagation based multilayer perceptron (MLP). The preliminary experiments show that such a system can classify human hand gestures with a classification accuracy of 97%.