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

Deep learning-based object detection research has primarily evolved around RGB camera imagery. While state-of-the-art models like YOLO, SSD, and RetinaNet demonstrate high accuracy in RGB images, their direct application to thermal imagery faces performance degradation due to the inherent differences between RGB and thermal characteristics, coupled with computational demands unsuitable for embedded environments. These challenges are particularly pronounced when using low-resolution thermal sensors, where object boundary ambiguity and thermal pattern uncertainty become significant obstacles. We propose FLARE (Fast and Lightweight Architecture for Real-time Estimation), an ultra-lightweight deep learning model for real-time object detection in low-resolution ( $32\times 24$ ) thermal images. FLARE implements a Feature Compression Block for efficient feature map compression and integration, along with a Spatial Denoise Block that enables adaptive processing across temperature regions to handle thermal sensor noise characteristics. To address the scarcity of thermal data, we developed the TOI (Thermal Object Insertion) data augmentation technique, which generates new training data while preserving thermal pattern characteristics. A composite loss function combining box, classification, and confidence losses was implemented to enhance object detection performance. Implementation on an STM32 MCU demonstrated that our proposed model achieved a higher mAP compared to existing lightweight models, while reducing Flash usage by 40.39 %, RAM by 49.02 %, inference time by 35.18 %, and MAC operations by 38.00 %. Comparative experiments with YOLOv8n, a state-of-the-art lightweight object detection model, on Raspberry Pi 4B showed that FLARE achieved higher mAP (0.72% improvement) while utilizing only 4.72% of YOLOv8n’s memory consumption and achieving 11.34 times faster execution speed.