Robust Activity Recognition via Redundancy-Aware CNNs and Novel Pooling for Noisy Mobile Sensor Data

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for mobile health and IoT applications.
• Existing methods often struggle with noise and feature discrimination in sensor data.
• Smartphone accelerometers offer rich data for HAR but require robust processing.

Purpose of the Study:

• To develop a robust Convolutional Neural Network (CNN) architecture for HAR using smartphone accelerometer data.
• To introduce and evaluate novel pooling mechanisms for enhanced feature discrimination and noise resilience.
• To investigate the efficacy of histogram-based image encoding for HAR.

Main Methods:

• Proposed two novel pooling mechanisms: Extrema Contrast Pooling (ECP) and Center Minus Variation (CMV).
• Implemented a 1D CNN on raw sensor streams and a 2D CNN with histogram-based image encoding.
• Normalized input data to [0, 1] for bounded pooled outputs.
• Conducted ablation studies to assess the contribution of different components.

Main Results:

• The 2D CNN system with histogram encoding achieved up to 96.84% weighted classification accuracy on the WISDM dataset.
• The proposed pooling layers (ECP and CMV) demonstrated consistent performance improvement and noise robustness.
• Histogram encoding provided the most significant performance enhancement, followed by the combination of ECP and CMV.
• The system showed reduced performance degradation under various noise conditions (Gaussian, salt-and-pepper, mixed).

Conclusions:

• The proposed CNN architecture with novel pooling and histogram encoding offers a robust and accurate solution for mobile HAR.
• Redundancy-aware pooling and histogram-based representations are beneficial for developing stable real-world HAR systems.
• The developed methods significantly outperform baseline models and traditional pooling techniques in noisy environments.