A hand gesture recognition system based on canonical superpixel-graph

workflow

Abstract

This paper presents a new hand gesture recognition system based on a novel canonical superpixel-graph earth mover’s distance (CSG-EMD) metric. It aims to improve the performance of the superpixel earth mover’s distance (SP-EMD), a recently proposed distance metric designed for depth-based hand gesture recognition. In real life, people have their own habits while performing certain hand gestures, which yields a variety of hand shapes with different finger poses. Such variety may affect the accuracy of SP-EMD and hence will degrade its performance. In this paper, we propose a new distance metric CSG-EMD to alleviate the problem. Scattered superpixels are organized in the form of canonical superpixel-graph which can factor out non-standard finger poses, resulting a well-structured finger-pose-neutral shape representation for hand gestures. Moreover, a structure stress based fusion scheme is applied to formulate the proposed distance metric, i.e. CSG-EMD, for gesture recognition. Experimental results on five public gesture datasets show that the proposed CSG-EMD-based system can achieve better recognition accuracy than other state-of-the-art algorithms compared. Its superiority is further demonstrated by two real-life applications.

Publication
Signal Processing: Image Communication

本论文获得2017-2018年度宁波市自然科学优秀论文奖 三等奖

Chong Wang
Chong Wang
Associate Professor

My research interests include hand gesture recognition, zero-shot learning, action recognition, image/video processing.