Gesture-based Intention Prediction for Automatic Door Opening using Low-Resolution Thermal Sensors: A U-Net-based Deep Learning Approach

Abstract

Personal health consciousness has increased amid pandemics. The implementation of automatic doors could help stop the infection. The need for an intelligent sensor emerges for automatic doors to prevent unneeded open as well as customer privacy concerns. This research proposes a novel automatic door opening mechanism using a low-resolution thermal sensor, based on which a multi-task U-Net structure network is adopted to classify hand-raising gestures. With the aid of segmentation masking, there is 74% reduction of training steps for convergence than that of mere thermal image classification while maintaining similar classification performance. On-site deployment of this approach via constantly collecting door-opening misclassification cases for model improvement will lead to practical success in the near future.