Radar-Based Human Activity Recognition In An Indoor Environment

Human Activity Recognition (HAR) has attracted significant attention due to its wide range of applications, including ambient assisted living, healthcare monitoring, and smart environments. Traditional wearable sensor-based systems often suffer from user discomfort, compliance issues, and limited robustness. Video-based approaches, while informative, raise privacy concerns and struggle in low-light conditions or at extended ranges. This thesis introduces a Frequency-Modulated Continuous Wave (FMCW) radar-based framework for both multi-class human activity recognition and fall detection, leveraging a 60 GHz radar and multi-dimensional feature maps—namely Range-Doppler, Range-Azimuth, and Range-Elevation. Unlike conventional methods that treat these maps as images, this work directly inputs the multi-dimensional data into learning models, preserving the spatial and temporal structure of the radar signals. Two datasets were collected in realistic indoor environments—bedroom and living room settings—capturing a diverse range of activities, including less-studied and challenging motions as well as fall scenarios. A complete radar signal processing pipeline was designed to transform raw radar data into structured, informative features. A range of machine learning and deep learning models, including Support Vector Machines (SVM), Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), Convolutional LSTM (ConvLSTM), and hybrid CNN+LSTM and ResNet+LSTM architectures, were evaluated in terms of classification accuracy and computational efficiency. The models were assessed using both cross-scene and leave-one-person-out validation schemes to investigate their generalization across environments and subjects. The results demonstrate that the proposed radar-based approach offers high recognition performance and strong generalization capabilities, achieving F1-scores of up to 98% for 4-class activity recognition on the bedroom dataset, 93.5% for fall detection using the living room dataset, and 76.1% for 6-class activity recognition on the living room dataset. These findings highlight the system's suitability for near real-time and privacy-aware applications in indoor settings.