[Paper Review] WSense: A Robust Feature Learning Module for Lightweight Human Activity Recognition

Ayokunle Olalekan Ige, Mohd Halim Mohd Noor|arXiv (Cornell University)|Mar 31, 2023

Context-Aware Activity Recognition Systems4 citations

TL;DR

WSense is a lightweight, parameter-efficient feature learning module that enhances human activity recognition (HAR) using wearable sensor data. It employs two 1D CNNs and global max pooling to extract robust, window-size-invariant features, outperforming baselines across diverse sliding window sizes with minimal model size on WISDM and PAMAP2 datasets.

ABSTRACT

In recent times, various modules such as squeeze-and-excitation, and others have been proposed to improve the quality of features learned from wearable sensor signals. However, these modules often cause the number of parameters to be large, which is not suitable for building lightweight human activity recognition models which can be easily deployed on end devices. In this research, we propose a feature learning module, termed WSense, which uses two 1D CNN and global max pooling layers to extract similar quality features from wearable sensor data while ignoring the difference in activity recognition models caused by the size of the sliding window. Experiments were carried out using CNN and ConvLSTM feature learning pipelines on a dataset obtained with a single accelerometer (WISDM) and another obtained using the fusion of accelerometers, gyroscopes, and magnetometers (PAMAP2) under various sliding window sizes. A total of nine hundred sixty (960) experiments were conducted to validate the WSense module against baselines and existing methods on the two datasets. The results showed that the WSense module aided pipelines in learning similar quality features and outperformed the baselines and existing models with a minimal and uniform model size across all sliding window segmentations. The code is available at https://github.com/AOige/WSense.

Motivation & Objective

• To address the high parameter count in existing attention-based feature learning modules like squeeze-and-excitation, which hinder deployment on resource-constrained edge devices.
• To develop a feature learning module that maintains high performance across varying sliding window sizes in wearable sensor-based HAR.
• To ensure consistent feature quality regardless of input window segmentation, improving model robustness.
• To reduce model complexity while maintaining or improving recognition accuracy on accelerometer and multimodal sensor data.
• To enable efficient deployment of HAR models on end devices through minimal and uniform model size.

Proposed method

• WSense uses two 1D convolutional layers to extract hierarchical temporal features from raw wearable sensor signals.
• Global max pooling is applied after each 1D convolutional layer to reduce spatial dimensions and retain salient features.
• The module is designed to be agnostic to the size of the input sliding window, ensuring consistent feature representation across different window segmentations.
• The architecture is integrated into both CNN and ConvLSTM pipelines for evaluation on multimodal and single-sensor datasets.
• The module is trained end-to-end with standard cross-entropy loss and optimized via stochastic gradient descent.
• The design emphasizes parameter efficiency, avoiding attention mechanisms or heavy normalization layers to minimize model size.

Experimental results

Research questions

• RQ1 Can a lightweight feature learning module maintain high performance across varying sliding window sizes in wearable sensor-based HAR?
• RQ2 Does WSense achieve comparable or better accuracy than existing attention-based modules while using significantly fewer parameters?
• RQ3 How does WSense perform when integrated into both CNN and ConvLSTM architectures on different sensor datasets?
• RQ4 Is the feature representation from WSense robust to changes in input window length and sensor modality?
• RQ5 Can WSense enable efficient deployment of HAR models on edge devices due to its minimal and uniform model size?

Key findings

• WSense achieved superior or competitive accuracy compared to baseline models across all 960 experiments conducted on WISDM and PAMAP2 datasets.
• The module maintained consistent performance across all tested sliding window sizes, demonstrating robustness to input segmentation variations.
• WSense significantly reduced model size compared to attention-based modules like squeeze-and-excitation, enabling efficient deployment on edge devices.
• On the PAMAP2 dataset, WSense outperformed existing methods in terms of accuracy while using fewer parameters.
• The integration of WSense into both CNN and ConvLSTM pipelines led to improved feature learning quality without increasing model complexity.
• The global max pooling mechanism effectively preserved discriminative features while reducing computational overhead.