Abnormal Activity Detection Using Pyroelectric Infrared Sensors.

Area of Science:

• Computer Science
• Artificial Intelligence
• Gerontology

Background:

• Healthy aging is a significant societal concern.
• Detecting abnormal activities in living environments is essential for elder care.
• Current methods often require laborious and inconsistent manual data labeling.

Purpose of the Study:

• To develop an automated method for abnormal activity detection.
• To eliminate the need for manually labeled training data.
• To leverage spatio-temporal characteristics of human activity using PIR sensors.

Main Methods:

• Utilizing Field of View (FOV) modulation to encode human activity into low-dimension data streams from PIR sensors.
• Measuring normal activity similarity using Kullback-Leibler (KL) divergence.
• Employing self-tuning spectral clustering for unsupervised discovery of normal activity clusters.
• Modeling normal activities with Hidden Markov Models (HMMs) and profiling with One-Class Support Vector Machines (OSVMs).

Main Results:

• The proposed method successfully detected abnormal activities in real indoor environments.
• The system demonstrated efficacy using only normal activity samples for training.
• The approach avoids the challenges associated with manual data annotation.

Conclusions:

• The developed unsupervised method offers an effective solution for abnormal activity detection in healthy aging contexts.
• This technique reduces the burden of data labeling, making it more practical for real-world applications.
• The findings support the use of PIR sensor data and advanced machine learning for monitoring and safety.