A 181µW Real-Time 3-D Hand Gesture Recognition System based on Bi-directional Convolution and Computing-Efficient Feature Clustering

Abstract

Vision-based hand gesture recognition (HGR) system, as an intuitive and portable approach for human-computer interaction (HCI), has been widely deployed on smart edge devices. While the prior endeavors remain different limitations to achieve a balance between power consumption and stability of the system. The HGR processors based on deep neural networks [1]–[3] achieved high recognition accuracy at the cost of significant power consumption. In contrast, the emerging energy-efficient HGR systems [4]–[5] based on ultra-compact customized algorithms suffer from performance degradation as the disturbing factors in the background increase.