A Deep Learning-Based Semantic Segmentation Model Using MCNN and Attention Layer for Human Activity Recognition.

Area of Science:

• Wearable technology
• Human activity recognition
• Deep learning for time-series analysis

Background:

• Wearable devices like smartwatches collect sensor data for human activity recognition.
• Traditional methods use fixed-size windows for time-series data, limiting performance due to unknown activity durations.
• Sliding window approaches attempt to mitigate this but still face challenges.

Purpose of the Study:

• To propose a novel human activity recognition method that classifies all data frames, overcoming limitations of fixed-window and sliding-window techniques.
• To enhance model performance using fused features from multiple Convolutional Neural Networks (CNNs) and an attention mechanism.
• To validate the proposed method's effectiveness through comprehensive evaluation experiments.

Main Methods:

• A frame-level classification approach is proposed, processing all data points rather than fixed or sliding windows.
• Features are extracted using multiple CNNs with varying kernel sizes and fused for comprehensive representation.
• An attention layer is incorporated at channel and spatial levels to refine feature importance and improve recognition.

Main Results:

• The proposed model achieved the highest F1-score (over 0.9) across all target activities compared to other deep learning models.
• The novel method demonstrated a significant improvement over sliding window (SW) methods, with a 0.154 higher F1-score.
• The designed model showed a substantial F1-score increase of 0.184 compared to baseline methods.

Conclusions:

• The proposed frame-level classification method significantly enhances human activity recognition accuracy.
• The fusion of multi-kernel CNN features and attention mechanisms is effective in improving model performance.
• This approach offers a more robust and accurate solution for activity recognition using wearable sensor data.