Real-Time Joint Classification and Heat-Source Localization on Native 8×8 Thermal Frames with a Lightweight Dual-Head CNN for Microcontroller Deployment

Abstract Indoor monitoring with RGB cameras records identifiable imagery, which is hard to justify in privacy-sensitive spaces. Ultra-low-resolution infrared arrays such as the Panasonic AMG8833 Grid-EYE avoid this at the hardware level: an 8×8 grid of 64 temperature values cannot identify a person, yet still carries enough structure to separate an empty room, an occupant, and a fire event. Prior work on this sensor class has relied on handcrafted pipelines that rarely address joint recognition and localization at native resolution. We present JFilterLocalizationCNN, a compact dual-head convolutional network (~115,365 parameters) that classifies the thermal scene into three categories and regresses the (x,y) position of the dominant heat source in one forward pass over the raw 8×8 frame. A lightweight J-filter removes the ambient offset through per-frame background subtraction; per-frame normalization stabilizes training without distorting the spatial signatures localization depends on. On 12,000 balanced frames, the model reaches 99.75% classification accuracy and 0.39-pixel mean absolute localization error. Each inference costs roughly 1.5 million multiply-accumulate operations and 2.7 ms on a desktop CPU, within the 100 ms budget of a 10 Hz sensor. Post-training INT8 quantization compresses the model from 466 KB to 126 KB with no measurable loss of accuracy or localization precision. The same INT8 TensorFlow Lite model runs on a Seeed XIAO ESP32-S3 with the AMG8833 and reaches 81.52% accuracy (macro-F1 = 0.82) and 1.39-pixel mean localization error on 4,752 live frames. Cross-dataset experiments and a twelve-case ablation further characterize generalization and the contribution of each design choice.