Efficiently improving the Wi-Fi-based human activity recognition, using auditory features, autoencoders, and...

Area of Science:

• * Computer Science
• * Signal Processing
• * Machine Learning

Background:

• * Human Activity Recognition (HAR) using Wi-Fi signals is gaining traction due to infrastructure availability.
• * Limited training data is a significant challenge for achieving high accuracy in HAR.
• * Channel State Information (CSI) reflects Wi-Fi signal propagation but requires substantial data for effective analysis.

Purpose of the Study:

• * To develop deep learning models for HAR that achieve high accuracy with significantly reduced training data.
• * To leverage pretrained encoders for efficient feature extraction in Wi-Fi-based HAR.
• * To evaluate the effectiveness of a pretrained encoder for improving HAR performance under data constraints.

Main Methods:

• * Utilized a pretrained encoder from a Multi-Input Multi-Output Autoencoder (MIMO AE) trained on Mel Frequency Cepstral Coefficients (MFCC).
• * Employed a fine-tuning strategy, incorporating the pretrained encoder as a fixed layer in a deep learning classifier.
• * Evaluated model performance using K-fold cross-validation (K=5) with limited training data (30% of available data).

Main Results:

• * The proposed method achieved 90.3% accuracy using only 30% of the training/validation data, a 17.7% improvement over a classifier without the encoder (79.3% accuracy).
• * Using the pretrained encoder as a fixed layer demonstrated significant efficiency, with minimal accuracy gain (2.4%) when treated as a trainable layer.
• * The approach effectively addresses data scarcity challenges in Wi-Fi-based HAR.

Conclusions:

• * A pretrained encoder significantly enhances HAR accuracy using Wi-Fi signals, even with limited training data.
• * The proposed fine-tuning method offers an efficient and effective solution for data-constrained HAR applications.
• * This technique shows great promise for practical HAR systems relying on Wi-Fi infrastructure.