Device-Free WiFi Sensing for Human Activity Recognition
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Human activity recognition (HAR) using WiFi signals (WiFi-based HAR) has drawn considerable interest from the research community. In contrast to traditional device-based sensing techniques, WiFi-based HAR possesses several advantages, including convenience, wide availability, and privacy protection, making it an attractive sensing solution for a wide range of applications in smart home, health care, and intelligent monitoring.
Recently, applying deep learning (DL) to WiFi-based HAR has received strong research interest. Assisted by signal processing techniques, DL-based HAR methods are able to automatically extract deep features from input signals, contributing to successful recognitions. Despite its effectiveness in improving recognition performance, DL-based HAR methods suffer from several inherent drawbacks. First, feature extraction is a challenging task that always bottlenecks the recognition performance. Second, DL-based HAR requires a large number of training examples from the testing/targeted environment or/and previously seen environments (PSEs) to train the corresponding DL architectures. When the number of required samples is not sufficient, the sensing performance will drop dramatically. Third, the trained model in one environment cannot be directly applied to another environment without additional effort.
My PhD thesis aims to provide novel solutions to the above WiFi-based HAR issues. Specifically, to extract effective features, we propose two advanced methods together with leveraging the property of DL architectures to enhance the quality of input signals of DL networks and extracted representative features. For a reliable recognition with limited training samples, we propose a novel HAR scheme by developing innovative signal processing methods and exploring the characteristics of one-shot learning to reduce the number of required training samples. The proposed HAR scheme is able to accomplish successful recognitions when both the number of PSEs and the amount of samples from the testing environment are quite limited (e.g., one PSE and at the minimum one sample for each activity from the testing environment). To achieve environmental robustness, we propose two novel signal processing algorithms and leverage the features of the matching network. The proposed models are trained once and can be directly applied to various new/testing environments for reliable recognitions without requiring an additional retraining process.
Please use this identifier to cite or link to this item: