Enhancing Radar-Based Human Activity Recognition: A Multi-Stream Deep Learning Approach with Advanced Signal Processing and Data Augmentation

Human Activity Recognition (HAR) using radar technology offers a privacy-preserving and robust alternative to traditional optical sensors, with applications in healthcare, assisted living, and security. This research develops a robust radar-based HAR system using Frequency-Modulated Continuous Wave (FMCW) radar and the Human Activity Radar Challenge Glasgow (HARCG) dataset. The study addresses challenges in signal processing, data scarcity, and model generalization by integrating advanced signal processing techniques, dimensionality reduction, synthetic data augmentation, and deep learning methodologies. The proposed multi-stage processing pipeline includes clutter and noise suppression, feature extraction through Short-Time Fourier Transform (STFT) and Range-Doppler analysis, and dimensionality reduction using generalized 2D Principal Component Analysis (2D-PCA). Deep learning architectures, including Residual Neural Networks (DivNet-15), Convolutional Autoencoders (CAE), and Long Short-Term Memory (LSTM) networks, are employed to capture spatiotemporal dependencies in radar signals. Ensemble learning techniques enhance classification robustness. Evaluated using the HARCG dataset, the system demonstrates significant improvements in activity classification accuracy, highlighting the potential of radar-based HAR for real-time, non-intrusive human activity monitoring in healthcare, security, and smart home automation.