Hybrid Spatial-Temporal Deep Learning Architectures for FMCW Radar-Based Human Activity Recognition
TL;DR
A hybrid spatial–temporal deep learning framework for FMCW radar-based HAR using micro-Doppler spectrograms is proposed, which outperform conventional CNN and standalone LSTM approaches in accuracy, robustness, and generalisation.
Abstract
Human Activity Recognition (HAR) supports intelligent healthcare, surveillance, assisted living, and human–machine interaction. Vision-based methods are often limited by privacy concerns, illumination changes, and occlusion. This study proposes a hybrid spatial–temporal deep learning framework for FMCW radar-based HAR using micro-Doppler spectrograms. Four architectures are compared: 3D CNN–LSTM, 3D Bi-LSTM–CNN, CNN–Dilated Convolution–LSTM, and a Hybrid Ensemble CNN-LSTM with a Decision Tree classifier. Radar processing includes beat-frequency extraction, Range FFT, Doppler FFT, clutter suppression, and spectrogram generation. Convolutional layers extract spatial features, while LSTM and Bi-LSTM networks model temporal dependencies; dilated convolution expands the receptive field efficiently. Experimental results show that the hybrid models outperform conventional CNN and standalone LSTM approaches in accuracy, robustness, and generalisation. The hybrid ensemble achieves the best performance by combining spatial–temporal learning with ensemble optimisation while remaining effective in noisy environments and preserving user privacy.