| dc.description.abstract | Embedded monitoring systems that rely on RGB cameras raise well documented privacy
concerns in sensitive spaces. Low-resolution thermal sensors offer an alternative: The Panasonic
AMG8833 Grid-EYE outputs only an 8×8 grid of 64 temperature values too coarse to identify
individuals, yet structured enough to distinguish an empty room, a person, and a fire event.
This thesis proposes JFilterLocalizationCNN, a lightweight dual-head convolutional neural net-
work for simultaneous classification and heat-source localization on native 8×8 thermal data.
Three classes are used - No Object, Object, and Object with Fire - with normalized (x, y) coordi-
nates for the dominant heat source. A J-filter removes ambient offset via per-frame background
subtraction; normalization maps readings to [0, 1] for stable training without distorting spatial
signatures needed for localization.
Training uses AdamW with combined cross-entropy and MSE losses on 3,087 labeled frames.
On a stratified test set of 618 samples, the model reaches 99.68% classification accuracy (F1
above 0.99 per class) and 0.37-pixel MAE localization (RMSE 0.67 pixels). The two tasks share
one forward pass and one set of convolutional weights, which keeps the design small enough to
reason about on paper and, in principle, to port to an embedded runtime later.
The experiments here were conducted in a single indoor setting with a controlled heat source
for the fire-related class. That scope limits how far one should generalize the numbers, yet they
still suggest that joint learning on native 8×8 frames is workable when preprocessing preserves
relative heat structure. On-device latency, power use, and post-training quantization are left for
future work.Chapters that follow spell out the dataset, preprocessing, layer stack, training schedule, and metrics in full. The aim is a thesis someone can audit line by line rather than a high-level claim alone. | en_US |
