Human Sensing Using Low Resolution Thermopile Sensors
Files
Date
Authors
Ma, Shengjun
Supervisor
Alam, Fakhrul
Konings, Daniel
Lai, Edmund
Item type
Thesis
Degree name
Journal Title
Journal ISSN
Volume Title
Publisher
Auckland University of Technology
Abstract
Interest in human sensing has grown substantially in recent years. This growth is driven by applications across diverse domains, including healthcare, residential aged care, security and surveillance, entertainment, and intelligent living environments. A fundamental requirement for many human sensing solutions is the use of device free methods that safeguard user privacy. Although various sensing modalities have been explored, each has its own advantages and limitations. Thermal sensing, which operates by detecting and measuring Infrared (IR) radiation emitted by the human body, is emerging as a promising option for device free human sensing. This thesis presents a device-free and privacy-preserving framework for human sensing applications using the IR signal captured by low resolution thermopile sensors. Machine learning approaches are applied to perform two distinct tasks: localization and fall detection. We develop a scalable, networked thermopile sensing platform to perform these tasks. To the best of our knowledge, no open-source or commercially available hardware platform has been specifically designed for human sensing using low-resolution thermopile sensors. We address this gap by releasing all PCB schematics, firmware, and supporting code. This enables replication and further extension by the wider research community. A 2D CNN–LSTM regression model was developed for localization and trained using data collected from eight participants. The model achieved median localization errors between 0.2 m and 0.3 m. The evaluation was performed on participants whose data were not used for training, ensuring robust and generalizable performance. We systematically examine multiple ceiling- and wall-mounted sensor configurations, ad-dressing limitations of prior studies that considered only a small number of sensors in fixed layouts. The results show a clear dependence of localization accuracy on both sen-sor number and placement. Experimental benchmarking against previous approaches indicates that the proposed algorithm provides improved localization accuracy. To assess scalability, four ceiling mounted sensors were deployed to cover a substantially larger area than reported in existing studies. The system operated reliably in this setting, demonstrating the practical viability of thermopile based localization. We developed a fall detection system that employs machine learning models to automatically identify falls from the IR data captured by the thermopile sensors. Three machine learning classifiers, namely SVM, MLP and RF, were trained on data collected from five participants enacting fall scenarios and performing other activities such as standing and sitting. The accuracy of the fall detection algorithms was evaluated using data from an additional five participants who were not included in the training set. The SVM classifier demonstrated the most consistent performance within the 16 different thermopile sensor configurations investigated. The accuracy varies between 92.5%and 99.2% and depends on the number of sensors and their respective locations. A combination of ceiling- and wall-mounted sensors with overlapping fields of view offered the best balance between coverage and sensor count. However, our experimental results also show that accurate fall detection is achievable with a single sensor, provided the fall occurs within the sensor’s clear field of view.Description
Keywords
Device-Free Localization, Fall Detection, Human Activity Recognition, Human Sensing, Indoor Localization, IR Sensing, Passive Localization, Thermal Sensor, Thermopile