Automatic lighting control based on person localization using Pseudo-CAM on edge devices

To improve energy efficiency and user comfort in smart office environments, we propose an automatic lighting control system that uses a ceiling-mounted camera and a lightweight CNN for person detection. Instead of conventional sensors, which often fail to detect stationary objects, our system leverages image-based inference and introduces a pseudo-class activation map (pseudo-CAM) for efficient localization without costly computation on resource constrained edge devices such as the Raspberry Pi Zero 2. Based on the estimated person’s location, illuminance is estimated locally to adjust LED luminance. Experiments using a 1/12 scale office model demonstrated accurate detection of stationary person objects and increased the frame rate by 2 times compared to a conventional CAM.