Contact-free Human Activity Sensing Using Wireless Signals
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Human activity sensing has been widely used in various fields such as security, autonomous driving, and human-computer interaction and has essential research significance and application value. Wireless signal-based activity sensing is based on the fact that human activities impact the wireless signal propagations such as reflection, diffraction and scattering, which provides human activity sensing opportunities through analyzing and mapping the variations to the received signals with a specific activity. Compared with other pathways, human sensing with wireless signals has unique advantages: device-free sensing pattern, robustness to environmental factors (e.g., weather, light, and temperature); the ability to penetrate obstacles; and protecting visual privacy.
WiFi and radar are commonly used wireless sensors for human sensing. In this dissertation, we first propose a channel state information (CSI)-based Doppler speed estimation method, which can provide accurate Doppler estimations with phase offset removal for further human activity analysis. However, using the estimated Doppler frequency estimations alone generally cannot obtain satisfactory performance for human activity recognition. By contrast, radar is a natural sensing sensor and can be utilized to estimate activity-related parameters (e.g., the time-varying range and Doppler frequency information) more easily than WiFi signals, we go a step beyond activity parameter estimation and focus on activity recognition with radar signals. Specifically, the main research problems and contributions of this thesis can be summarized as follows.
First, to remove the carrier frequency offset caused by clock asynchronism and attain accurate Doppler speed estimates, we study Doppler frequency estimation using the cross-antenna signal ratio (CASR) method for scenarios with general movement. Using a publically available ii WiFi CSI dataset Widar 2.0, we then validate the efficiency of the proposed Doppler frequency estimation algorithms.
Second, aiming at enhancing the generalization ability of deep learning (DL) methods to human individual differences and improving the HAR performance on different persons’ activities, we present an instance-based transfer learning approach ITL for cross-target HAR with radar spectrograms. Experiments demonstrate that the proposed approach is more generalized to the data distribution discrepancy and can scale well to recognize different persons’ activities.
Third, we propose a supervised few-shot adversarial domain adaptation (FS-ADA) method for radar-based HAR. This method does not require a large number of radar data for training when applied to a new environment. Experimental results on two few-shot radar-based HAR tasks show that the proposed FS-ADA method is effective for few-shot HAR, and outperforms state-of-the-art methods.
Please use this identifier to cite or link to this item: