Effects of sliding window variation in the performance of acceleration-based human activity recognition using deep...

Area of Science:

• Computer Science
• Biomedical Engineering

Background:

• Deep learning (DL) models significantly enhance human activity recognition (HAR) accuracy compared to traditional methods.
• DL models extract features from raw time-series acceleration data, improving HAR performance.
• Sliding windows are a key feature extraction technique for time-series data preprocessing.

Purpose of the Study:

• To determine the optimal sliding window size for HAR using DL models.
• To analyze the impact of window size on accuracy and processing efficiency.
• To compare model performance across different window sizes and data sources (IMU and MOCAP).

Main Methods:

• Four DL models were evaluated: DNN, CNN, LSTM, and CNN-LSTM.
• Sliding window sizes ranged from 5 to 200 frames with fixed overlap.
• Performance was assessed using raw acceleration data from wearable IMU sensors and motion caption systems (MOCAP) for walking, sit-to-stand, and squatting activities.

Main Results:

• An optimal window size of 20-25 frames (0.20-0.25s) was identified for both IMU and MOCAP data, achieving high accuracy (up to 99.07%) and F1-scores (up to 87.08% with CNN-LSTM).
• Larger window sizes (100-200 frames) showed minimal accuracy gains but reduced processing efficiency.
• Smaller windows (below 20 frames) resulted in a decrease in F1-score, indicating a loss of performance.

Conclusions:

• The 20-25 frame sliding window offers the best trade-off between accuracy and processing speed for HAR.
• Deep learning models, particularly CNN-LSTM and LSTM, are effective for HAR with optimized window sizes.
• Efficient preprocessing via optimized sliding windows is crucial for real-time HAR applications.