Stacked LSTM for Human Activity Recognition via Radar Sensing

Recent advances on Radar-based human activity recognition (HAR) have revealed the potential of this technique in a wide range of application scenarios, including security and surveillance as well as non-intrusive monitoring. In this paper, we propose a two-layer stacked Long Short-Term Memory (LSTM) model combined with a Spectrogram Binning technique for human activity recognition using continuous wave radar signals. The primary contribution of this study lies in the spectrogram binning technique, which effectively reduces the dimensionality of the input feature vector, minimizing computational complexity without sacrificing crucial temporal information. This reduced feature vector is then fed into the proposed stacked LSTM model, which leverages its hierarchical structure to capture temporal dependencies and time-varying Doppler and micro-Doppler signatures essential for characterizing human motion. By combining these innovations, the proposed approach was demonstrated effectively on the DIAT-μRadHAR dataset, a diverse experimental dataset developed by DIAT, Pune with a high model accuracy of 98.10% with lower computational costs and model complexity.