Reproduce “Person Detection Using an Ultra Low-resolution Thermal Imager on a Low-cost MCU” on OpenResearch

About this paper

Maarten Vandersteegen, Wouter Reusen, Kristof Van Beeck, Toon Goedemé

Researchers developed an ultra-lightweight convolutional neural network for real-time person detection using an ultra-low-resolution (32x24 pixel) thermal imager on low-cost microcontrollers. The system combines a CNN with a novel background subtraction mechanism and aggressive model compression, achieving F1-scores of 79.9-91.62% and outperforming existing commercial solutions by up to 15.6% while operating within stringent memory and power constraints.

Key insights

• The integration of a classical background subtraction technique significantly boosts person detection accuracy (21-28% AP increase) on ultra low-resolution thermal data by providing crucial temporal context.
• Extreme model compression (up to 136x parameter reduction) can be achieved with minimal accuracy loss (less than 1% AP drop for 90-degree view, ~4% for 45-degree view) for constrained object detection tasks on thermal images.
• TinyML frameworks like microTVM and TFLite Micro enable real-time CNN inference (46-87ms per frame) for object detection on low-cost microcontrollers with minimal RAM (30.7kB) and flash memory (33.6kB).

Summary by alphaXiv.