Multi-Person Localization Based on a Thermopile Array Sensor with Machine Learning and a Generative Data Model.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Technology

Background:

• Thermopile sensor arrays offer a privacy-preserving method for person detection and localization in smart building automation.
• Machine learning algorithms like YOLOv5 and DETR are prominent for general object detection.

Purpose of the Study:

• To adapt YOLOv5 and DETR for localizing people in low-resolution (32x32 pixel) thermal array images.
• To evaluate the robustness and performance of these algorithms, enhanced by a novel image generation technique.

Main Methods:

• Adaptation of YOLOv5 and DETR algorithms for person localization in thermal images.
• Development of an image interpolation (IIG) generator to create synthetic thermal frames from sparse data.
• Conducting multiple robustness tests against various environmental and object-related disturbances.

Main Results:

• Both YOLOv5 and DETR achieved high mean average precision (mAP) exceeding 98%.
• Algorithms demonstrated robustness against warm air, sun radiation, sensor replacement, and various movement scenarios.
• Precision decreased with thick clothing and when the number of people exceeded training data.

Conclusions:

• Both YOLOv5 and DETR are suitable for person detection and localization using thermopile sensors.
• YOLOv5m offers advantages in real-time processing, smaller model size, and slightly higher precision.
• The novel image generator effectively counteracts precision loss due to sparse labeled data.