Real-Time Object Detection Using Low-Resolution Thermal Camera for Smart Ventilation Systems

We propose an ultra-lightweight deep learning model for real-time object detection in smart ventilation systems using low-resolution thermal imaging sensors. The proposed FLARE (Fast and Lightweight Architecture for Real-time Estimation) model employs a Spatial Denoise Block, TOI (Thermal Object Insertion) data augmentation, and a composite loss function. Implemented on an STM32 MCU,FLARE showed an 84.28% performance improvement, 53.37% speed increase, and over 40% Flash and RAM reduction compared to existing lightweight methods. On Raspberry Pi 4B, it achieved higher mAP (0.72% improvement) than YOLOv8n, with 1/118 memory usage and 12.05 times faster execution speed.