A Stacked Human Activity Recognition Model Based on Parallel Recurrent Network and Time Series Evidence Theory

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Accurate real-time human activity recognition is crucial for intelligent living conditions and health status monitoring.
• Posture analysis relies heavily on precise activity recognition from sensor data.
• Existing methods may face challenges with raw time-series data from wearable sensors.

Purpose of the Study:

• To propose a fine-grained evidence reasoning approach for reliable human activity recognition using raw time-series data.
• To optimize the selection of the basic time unit for balancing accuracy and computational cost.
• To enhance classification accuracy by reducing uncertainty through trainable evidence combination and inference.

Main Methods:

• Utilizing Long Short-Term Memory (LSTM) networks for feature extraction from multidimensional time-series data.
• Projecting raw sensor data into probability assignments.
• Implementing a trainable evidence combination and inference network for improved classification.

Main Results:

• The proposed approach demonstrates the effectiveness of fine granularity and evidence reasoning.
• Achieved a high recognition accuracy of 96.4% for human activity.
• Validated the method's effectiveness without introducing additional training complexity.

Conclusions:

• The fine-grained evidence reasoning approach offers a timely and reliable solution for human activity recognition.
• This method significantly improves posture analysis and has potential applications in health monitoring.
• The approach effectively handles raw time-series data from wearable sensors.