Multiple-Image Super-Resolution for Networked Extremely Low-Resolution Thermal Sensor Array

Observing activeness of patients and elderly for long time at private space is a desired metric for elderly care. Many state-of-the-art mechanisms require either using wearable sensors, assists by medical staffs, or observation at designated location. Consequently, it is difficult, if not impossible, to acquire long term trend. Thermal sensors are not sensitive to illumination and can detect heat sources under low lightness. Moreover, it is not trivial to identify an individual from thermal images. However, a low cost and effective solution remain open. Low-resolution thermal sensors cost less but suffer from noisy reading. In this work, we fuse multiple low-resolution thermal images to reconstruct higher resolution thermal images to identify human activities. The evaluation results show that the accuracy of activity recognition while using low resolution thermal images are 96% of that while using high resolution thermal images, which requires 80 times of data and ten times of hardware cost.