Privacy-Preserving Real-Time Bed Occupancy and Posture Monitoring Using Low-Resolution Thermal Array

LIU Xuan: 3035973799

Supervisor: Prof. Wu Chenshu

This project produces an affordable, privacy-preserving, non-contact sleep monitoring solution leveraging thermal array (IRA) sensors, a low-cost, human-sensitive modality that enables passive monitoring without ambient illumination.

Code open-sourced on GitHub Page

Introduction

1. Sleep is a vital physiological indicator. Monitoring sleep enables early detection of many abnormalities that are otherwise difficult to perceive.
2. Monitoring bed-side activities also enable real-time warning of activities related to fall risks, such as bed exit events.
3. Clinical monitoring solutions require specialized personnel, venues, and equipment, making it unsuitable for long-term, continuous self-monitoring at home.

Therefore, we hope to provide a novel solution that is affordable, non-intrusive, privacy-preserving, and human-sensitive, to transparently monitor sleep and report relevant statistics.

Methods

Sensor Choice

We select a low-cost thermal array (IRA) sensor that produces heat maps, IRA properties:

1. Sensitve to human presence by detecting the thermal signals .
2. Functional in complete darkness.
3. Privacy preserving due to low-resolution.
4. Relatively low-cost.

Which make it a suitable modality for home-based monitoring.

System Design Overview

To effectively monitor human presence and posture on bed, we designed a system with six modules for thermal map collection, processing, analysis and statistics output.

Challenges

We identified three major challenges for in-bed monitoring using thermal arrays.

1. Bedding occlusion of thermal signals.
2. Thermal residual interference.
3. Low-resolution arrays hindering detailed feature extraction.

(A). Bedding occlusion; (B). Heat residual resembles a human. Low resolution and lack of feature makes distinction difficult.

System Components

• Data collection module: collect data from IRA sensor.
• Heat patch detection module: preprocess thermal map and extract separate heat sources. We used adaptive Otsu threshold, erosion, and connected components analysis for clean heat source detection.
• Heat residual detection module: determine if a heat source is human or thermal interference, according to the heat patch temperature history.
• Heat patch tracker module: track the dynamics and histories of heat sources.
• Posture classification module: classify postures based on thermal map.
• Integration module: integrate heat source information and posture classification result to produce predicted posture sequence and issue bed exit alarms.

Results

Presence detection achieves over 95% accuracy. Posture detection achieves over 89% accuracy, and overall accuracy exceeds 89% for all held-out users and environments. This demonstrates the system’s robustness and effectiveness.

Cross-environment and cross-user performance are robust for the presence detection module, posture classification module, and overall system.

One-night sleep monitoring result shows some errors due to more diverse real-life poses, but still achieved 91.51% accuracy.

Case study: Exit detection using the system achieves an average latency of 1.09 seconds, and a detection rate of 98.3%

Contributions

1. A robust bed monitoring system using a low-cost, privacy-preserving thermal array.
2. Rule-based presence detection that enables accurate presence detection.
3. Validation of the sleep monitoring system showing promising cross-user and cross-environment generalizability.