Static Gesture Recognition Based on Thermal Imaging Sensors

Abstract

Gesture recognition technology has demonstrated its applicability in various scenarios within the field of human-computer interaction, owing to its flexibility and non-contact nature. In particular, thermal imaging technology which is not limited by lighting conditions and effectively reduces the risk of privacy disclosure by capture only thermal radiation rather than detailed features. In this study, a low resolution 32 x 24 pixels end-to-end embedded infrared thermal imaging camera gesture recognition system is developed. A thermal imaging gesture dataset of 4,500 images is constructed to train and evaluate the system. The effects on improve gesture recognition accuracy by research the spatial transform network attention mechanism (STN) is discussed. Thus, a new method combines lightweight Convolutional Neural Networks (CNN) and STN is propose. Additionally, the proposed method achieves 98.5% recognition accuracy and inference time of only 59 milliseconds per frame on embedded devices when tested on self-made infrared thermal imaging sign language gesture dataset, outperforming mainstream lightweight models.