CSI-Based Human Activity Recognition Using Multi-Input Multi-Output Autoencoder and Fine-Tuning.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Wi-Fi-based human activity recognition (HAR) is gaining traction due to accessible infrastructure.
• Channel State Information (CSI) offers higher accuracy for HAR than Received Signal Strength Indicator (RSSI).
• Limited training data is a significant challenge in many HAR applications.

Purpose of the Study:

• To develop deep learning models for accurate HAR using reduced training data.
• To investigate the effectiveness of pretrained encoders for feature extraction in HAR.
• To evaluate the performance improvement achieved through fine-tuning pretrained models.

Main Methods:

• Utilized a pretrained encoder for feature extraction from Wi-Fi CSI data.
• Employed fine-tuning techniques to integrate the encoder with a classifier.
• Trained deep learning models on a fraction of available data.
• Compared performance against models without a pretrained encoder.

Main Results:

• Achieved a 20% accuracy improvement using only 50% of the training data with a fine-tuned pretrained encoder.
• Demonstrated an 11% accuracy improvement with 50% of the training data using an untrainable encoder.
• Showcased the efficacy of deep learning with pretrained encoders for data-efficient HAR.

Conclusions:

• Pretrained encoders significantly enhance HAR accuracy, especially under data constraints.
• Fine-tuning pretrained models offers a robust approach to HAR with limited datasets.
• This method provides a practical solution for deploying HAR systems efficiently.