9.7 A 184 µ W Real-Time Hand-Gesture Recognition System with Hybrid Tiny Classifiers for Smart Wearable Devices
Yuncheng Lu, Van Loi Le, Tony Tae-Hyoung Kim · 2021
Recently, vision-based hand gesture recognition (HGR) has emerged as a natural and flexible human-computer interaction (HCI) approach. Users can control smart devices by applying hand gestures to imagers. However, prior efforts suffer from various limitations, such as excessive power consumption, low accuracy, and poor flexibility. The 3D HGR processors -[2] suffer from extremely large power overhead due to the employment of complex image processing, for example using Convolutional Neural Networks (CNNs). The grayscale sensor-based SoC [3] consumes less power. However, its accuracy is compromised, especially when the contrast between a hand gesture and the background is low. The infrared sensor-based SoC [4] can recognize 8 dynamic gestures with high accuracy (96%). Nevertheless, the over-simplified algorithm requires hand motion with a fixed gesture type, which limits the number of recognized dynamic gestures. Therefore, an ultra-low-power, flexible, and high accuracy HGR system is required.