Abstract: Device-free activity recognition and monitoring has become a promising research area with increasing public interest in pattern of life monitoring and chronic health conditions. This paper proposes a novel framework for in-home Wi-Fi signal-based activity recognition in e-healthcare applications using passive micro-Doppler (m-D) signature classification. The framework includes signal modeling, Doppler extraction, and m-D classification. A data collection campaign was designed to verify the framework where six m-D signatures corresponding to typical daily activities are successfully detected and classified using our software-defined radio (SDR) demo system. Analysis of the data focused on potential discriminative characteristics, such as maximum Doppler frequency and time duration of activity. Finally, a sparsity-induced classifier is applied for adapting the method in healthcare application scenarios and the results are compared with those from the well-known Support Vector Machine (SVM) method.
Introduction: Addressing the Challenges of Ambient Assisted Living
As the global population ages, monitoring the elderly or disabled in their homes—often referred to as Ambient Assisted Living—is becoming critical. Traditional monitoring relies on wearables, which users often forget or refuse to wear; cameras, which violate privacy; or PIR/RFID sensors, which offer low resolution. This research addresses the need for an unobtrusive, device-free way to reliably detect specific human movements, especially life-threatening ones like falls, in real-time. Crucially, healthcare applications require "one-shot" classification; the system must recognize an activity from a single motion cycle to trigger an instant alert, rather than relying on continuous, repeated movements.
System Architecture and Experimental Setup
The authors developed a passive Wi-Fi radar system using Software Defined Radios (SDRs) that leverages existing in-home Wi-Fi access points as transmitters of opportunity. The experimental setup, shown in Figure 1, utilized three synchronized USRP N210 SDRs—one for a reference channel and two for surveillance channels—to monitor a standard living room environment (3.87m x 3.67m).
Four male subjects performed six specific activities at a fixed location: picking an object up, sitting down, standing up from a chair, falling down, standing up after a fall, and getting out of bed. The resulting dataset consists of 170 motion samples per receiver channel.
Methodology: Signal Processing and Classification
The recognition pipeline consists of three main steps:
- Signal Extraction: The Cross Ambiguity Function (CAF) is used to extract micro-Doppler (m-D) signatures (time-velocity patterns) from Wi-Fi signals reflecting off the human body.
- Data Alignment: To account for varying motion durations, the system uses an automatic start/end point detection method based on the weighted standard deviation of frequency bins. Bi-cubic interpolation then resizes these signatures into uniform matrices.
- Sparse Representation Classification (SRC): Instead of manual feature extraction, the system applies Principal Component Analysis (PCA) for dimensionality reduction, followed by an SRC. The SRC uses a Subspace Pursuit algorithm to find the sparsest linear combination of training samples that can reconstruct a test sample, naturally ignoring misleading boundary samples between similar classes.
Results and Key Insights
The system proved highly effective at distinguishing between complex motions. A key insight was "embracing interference": while traditional radar filters out Direct Signal Interference (DSI) and ambiguity peaks, this system uses the presence or absence of a zero-Doppler line (caused by the body blocking the line-of-sight) as a discriminative feature.
As shown in Figure 2, the m-D signatures vary significantly between activities and receiver channels. The SRC achieved an average classification accuracy of 90.2% on Channel 1 and 85.2% on Channel 2, vastly outperforming the baseline SVM classifier, which only reached 46% accuracy. The results also highlighted sensitivity to receiver geometry; for instance, Channel 2 (bistatic) was superior at detecting falls compared to Channel 1 (quasi-monostatic) due to the stronger Doppler frequency induced relative to its position.
Conclusion and Future Work
While the current system demonstrates high accuracy, it is dependent on the physical angle and position of the receiver. Future work will focus on "joint sparsity-based classifiers" to fuse data from multiple Wi-Fi receiver channels simultaneously. By integrating multiple modalities, the system can overcome the blind spots of single receivers, increasing coverage and robustness for real-world healthcare applications.
Summary updated