Static Human Detection and Scenario Recognition via Wearable Thermal Sensing System

Qingquan Sun, Ju Shen, Haiyan Qiao, Xin-Lin Huang, Chen Chen, Fei Hu

TL;DR

This paper introduces a wearable thermal sensing system using pyroelectric infrared (PIR) sensors to detect static human presence and recognize indoor scenarios. By enabling active sensing through user movement, the system overcomes PIR sensors' traditional limitation of only detecting motion, offering a novel approach for environmental perception, particularly for visually impaired individuals.

What Problem Does This Paper Address?

Traditional wearable sensors primarily focus on the wearer's own physiological data and activity. However, they often fail to provide crucial information about the surrounding environment. This is a significant gap, especially for individuals with visual impairments, who need to understand their surroundings to navigate safely and avoid collisions. While existing human detection systems often rely on cameras, these are computationally intensive, data-heavy, and susceptible to lighting conditions. PIR sensors offer a low-cost, robust alternative, but they typically require movement to detect heat signatures, making them ineffective for static individuals. This paper tackles the challenge of using PIR sensors for static human detection and broader scenario recognition in indoor environments.

Methods

The core of the proposed system is a wearable thermal sensing module incorporating two pyroelectric infrared (PIR) sensors. To overcome the limitation of PIR sensors only detecting motion, the system employs an "active sensing" protocol. This means the user actively moves the wearable sensor (e.g., by swinging their arms) to create relative motion between the sensor and static targets, thereby generating detection signals.

Key technical contributions include:

  • Wearable PIR Sensor Node: A compact, low-power node featuring two PIR sensors arranged vertically to capture spatial information. It includes a conditioning circuit with a low-noise amplifier and a programmable system-on-chip (PSoC) for signal amplification and binary signal generation.
  • Binary Sampling and Analysis: A low-data-throughput scheme that converts analog PIR signals into binary streams. This simplifies processing and reduces transmission errors.
  • Active Sensing Protocol: The user's movement of the sensor node is crucial for detecting static targets. Different arm movements (left, right, up, down) are used to gather information about the target's size, shape, and orientation.
  • Statistical Feature Extraction: Three types of statistical features are extracted from the binary data:
  • Intersection Probability ($P_1$): Based on Buffon's Needle model, this feature quantifies the likelihood of a detection event, influenced by target size and sensor movement.
  • Temporal Statistical Feature ($\rho$): This measures the correlation between consecutive binary samples, reflecting the target's movement or posture changes.
  • Spatial Statistical Feature ($S_{jac}$): Using the Jaccard coefficient, this feature captures the geometric information from the two vertically aligned sensors, indicating similarity in spatial patterns.
  • Classifiers: k-centroid and k-nearest neighbors (knn) classifiers are used for recognition tasks.

Datasets

The paper does not explicitly mention pre-existing datasets. Instead, the authors describe their own experimental setup and data collection process. They created scenarios with static human subjects (standing and sitting) and non-human thermal sources (electronic devices) within a controlled 2.5m x 2.5m indoor area. The data collected consists of binary signals generated by the wearable PIR sensor node during these experimental conditions.

Key Insights and Ideas

  • Active Sensing for Static Detection: The most significant insight is the successful adaptation of PIR sensors for static target detection through user-initiated relative motion. This fundamentally changes how PIR sensors can be utilized for environmental awareness.
  • Binary Data Sufficiency: The paper demonstrates that even with low-data-throughput binary signals, rich information for human detection and scenario recognition can be extracted using appropriate statistical features.
  • Multi-Feature Fusion: Combining temporal, spatial, and intersection probability features provides a robust basis for distinguishing between different human statuses (standing/sitting) and even non-human thermal sources.
  • Potential for Assistive Technology: The system's portability and ability to provide environmental context make it a promising tool for assisting visually impaired individuals, offering them greater independence and safety.

Key Results

  • Static Human Detection: The system achieved high accuracy in detecting static human subjects.
  • Status Recognition: The knn classifier achieved over 90% accuracy in distinguishing between standing and sitting human subjects.
  • Non-Human Thermal Source Detection: Electronic devices were detected with high accuracy (up to 100%), with their distinct characteristics (regular shapes, different radiation wavelengths) leading to high temporal and spatial correlations.
  • Human Scenario Recognition: The system demonstrated good performance in recognizing scenarios with varying numbers of human subjects. For instance, scenarios with three or four subjects achieved accuracies of 87.22% (k-centroid) and 92.22% (knn).
  • Occlusion Challenge: The paper highlights that scenarios with occluded subjects (two subjects appearing as one) resulted in lower recognition accuracy, indicating a key area for future improvement.

Limitations and Future Work

  • Occlusion Handling: The authors acknowledge that occlusion remains a significant challenge, leading to lower accuracy in scenarios where subjects are partially hidden from the sensor's view. Improving detection in such cases is a primary direction for future work.
  • Complexity of Scenarios: While the system can recognize the number of humans, it currently focuses less on human interactions. Future work could explore more complex scenario recognition involving interactions.
  • Integration with Other Sensors: The authors suggest combining the wearable PIR sensor with other wearable sensors to expand its capabilities and applications.
  • Efficiency of Detection Method: Further improvements to the detection method, particularly for occlusion cases, are planned.

Summary updated

Abstract

Conventional wearable sensors are mainly used to detect the physiological and activity information of individuals who wear them, but fail to perceive the information of the surrounding environment. This paper presents a wearable thermal sensing system to detect and perceive the information of surrounding human subjects. The proposed system is developed based on a pyroelectric infrared sensor. Such a sensor system aims to provide surrounding information to blind people and people with weak visual capability to help them adapt to the environment and avoid collision. In order to achieve this goal, a low-cost, low-data-throughput binary sampling and analyzing scheme is proposed. We also developed a conditioning sensing circuit with a low-noise signal amplifier and programmable system on chip (PSoC) to adjust the amplification gain. Three statistical features in information space are extracted to recognize static humans and human scenarios in indoor environments. The results demonstrate that the proposed wearable thermal sensing system and binary statistical analysis method are efficient in static human detection and human scenario perception.
BibTeX
@inproceedings{sun2017static,
  author = {Qingquan Sun and Ju Shen and Haiyan Qiao and Xin-Lin Huang and Chen Chen and Fei Hu},
  title = {{Static Human Detection and Scenario Recognition via Wearable Thermal Sensing System}},
  booktitle = {Comput},
  year = {2017},
  url = {https://openreview.net/forum?id=YcjQbQA6qcL},
}
Figure