Cleaned Meta Pseudo Labels-Based Pet Behavior Recognition Using Time-Series Sensor Data

Area of Science:

• Animal Behavior
• Machine Learning
• Wearable Technology

Background:

• Increasing pet ownership necessitates reliable pet behavior recognition.
• Existing data collection methods face cost and reliability challenges.
• Sensor data from wearable devices offers a promising avenue for monitoring.

Purpose of the Study:

• To propose a novel method for classifying pet behavior using cleaned meta pseudo labels.
• To address limitations of conventional supervised learning and existing meta pseudo label techniques.
• To improve the accuracy and efficiency of pet behavior classification with limited labeled data.

Main Methods:

• Collected sensor data (accelerometers, gyroscopes, magnetometers) from wearable devices on pets.
• Classified five distinct pet behaviors.
• Developed and applied a cleaned meta pseudo label method, incorporating Distance Loss to reduce noise in unlabeled data.

Main Results:

• The proposed cleaned meta pseudo labels method achieved 88.3% accuracy.
• This significantly outperformed conventional supervised learning (82.9%) and existing meta pseudo labels (86.2%).
• The method demonstrated effectiveness in improving learning processes by noise removal from unlabeled data.

Conclusions:

• The cleaned meta pseudo label method offers a robust solution for pet behavior classification.
• This approach is particularly effective in scenarios with insufficient labeled data.
• Findings have significant implications for developing advanced pet monitoring systems.