Local frame match distance: A novel approach for exemplar gesture recognition

Abstract

Abstract-Gesture recognition using a training set of limited size for a large vocabulary of gestures is a challenging problem in computer vision. With few examples per gesture class, researchers often employ state-of-the-art exemplar-based methods such as Dynamic Time Warping (DTW). This paper makes two contributions in the area of exemplar-based gesture recognition. As an alternative to DTW, we first introduce the Local Frame Match Distance (LFMD), a novel approach for matching gestures inspired by a distance measure for strings, namely Local Rank Distance (LRD). While LRD efficiently approximates the non-alignment of character n-grams between two strings, we employ LFMD to efficiently measure the nonalignment of hand locations between two video sequences. Second of all, we transform LFMD into a kernel and use it in combination with Kernel Discriminant Analysis for sign language recognition with exemplars. The empirical results indicate that our method can generally yield better performance than a state-of-the-art DTW approach on the challenging task of American Sign Language recognition, while reducing the computational time by 30%.