Abstract: In this paper we propose the combination of graph-based characteristics and appearance-based descriptors such as detected edges for modeling static gestures. Initially we convolve the original image with a Gaussian kernel and blur the image. Canny edges are then extracted. The blurring is performed in order to enhance some characteristics in the image that are crucial for the topology of the gesture (especially when the fingers are overlapping). There are a large number of properties that can describe a graph, one of which is the adjacency matrix that describes the topology of the graph itself. We approximate the topology of the hand utilizing Neural Gas with Competitive Hebbian Learning, generating a graph. From the graph we extract the Laplacian matrix and calculate its spectrum. Both canny edges and Laplacian spectrum are used as features. As a classifier we employ Linear Discriminant Analysis with Bayes’ Rule. We apply our method on a published American Sign Language dataset (ten classes) and the results are very promising and further study of this approach is imminent from the authors.
Keywords: Gesture Recognition; Neural Gas; Linear Discriminant Analysis; Bayes Rule; Laplacian Matrix