A Semi-Supervised Transfer Learning with Dynamic Associate Domain Adaptation for Human Activity Recognition Using...

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Equipment-free human activity recognition is crucial for smart home applications.
• Existing methods often rely on wearable devices, limiting user freedom.
• Channel State Information (CSI) from WiFi signals offers a promising alternative for sensing human activities.

Purpose of the Study:

• To propose a semi-supervised transfer learning approach for environment-independent human activity recognition using WiFi CSI.
• To introduce a novel dynamic associate domain adaptation (DADA) method to address domain shift challenges.
• To develop an attention-based DenseNet (AD) model for enhanced feature extraction and recognition.

Main Methods:

• Data pre-processing techniques including missing packet filling, noise removal, and data augmentation were applied to improve CSI quality.
• A pre-trained model was developed on a labeled source domain dataset.
• The proposed dynamic associate domain adaptation with attention-based DenseNet (DADA-AD) was employed for semi-supervised transfer learning in target domains.

Main Results:

• The DADA-AD scheme achieved a high accuracy of 97.4% for human activity recognition across different domains.
• The proposed DADA method dynamically adjusts the ratio of labeled to unlabeled data, outperforming existing associate domain adaptation algorithms.
• Experimental results demonstrate the superiority of DADA-AD over current semi-supervised learning schemes in domain adaptation scenarios.

Conclusions:

• The DADA-AD approach effectively enables environment-independent human activity recognition using WiFi CSI.
• The dynamic adaptation strategy of DADA significantly mitigates the impact of environmental variations.
• This research offers a robust and accurate solution for advanced smart home applications without the need for wearable sensors.