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

This paper proposes a robust convolutional neural network (CNN) architecture for human activity recognition (HAR) using smartphone accelerometer data, evaluated on the WISDM dataset. We introduce two novel pooling mechanisms—Pooling A (Extrema Contrast Pooling (ECP)) and Pooling B (Center Minus Variation (CMV))—that enhance feature discrimination and noise robustness. Pooling A applies a nonlinear penalty based on the squared range between extrema, capturing sharp signal transitions, while Pooling B subtracts the standard deviation to penalize variability more smoothly, making it resilient to noise. To support these operations, input data are normalized to the [0,1] range, ensuring bounded, interpretable pooled outputs. Our dual-path framework includes: (1) a 1D CNN applied to raw tri-axial sensor streams with proposed pooling layers, and (2) a histogram-based image encoding pipeline that transforms segment-level sensor redundancy into RGB representations for a 2D CNN with fully connected layers. Across six activity classes, our system achieves up to 96.5% weighted classification accuracy, outperforming traditional average pooling and baseline models. Additionally, under Gaussian, salt-and-pepper, and mixed noise conditions, the proposed pooling layers consistently reduce performance degradation, demonstrating improved stability in real-world sensing environments. These results highlight the benefits of domain-informed pooling and data representations for mobile HAR systems, and offer a foundation for future IoT-based recommendation systems.