Human Motion Recognition by Micro-Doppler features and Concatenated CNN-LSTM Network

Human motion recognition (HMR) technology has emerged as a pivotal area of research within indoor human sensing. Traditional approaches to HMR often rely on uncomfortable wearable devices or visual systems with privacy concerns. Radar technology, with its noncontact features, respect for privacy, and capability to perform under any lighting conditions, presents a better alternative for HMR tasks. In this article, we utilize frequency-modulated continuous wave (FMCW) radar to collect echo signals from six types of human daily motions under indoor conditions and apply deep learning (DL) techniques to classify these motions based on the acquired radar data. Based on the characteristics of FMCW radar, we adopt a unique adaptive average threshold (AAT) filter to suppress stationary clutter and propose a DL model concatenated CNN-LSTM network (CCLN) based on concatenated 2-D convolutional neural networks (2DCNNs) and long short-term memory (LSTM) network. The experimental results demonstrate that the CCLN model achieved an accuracy rate of 95.91% in motion recognition, highlighting the effectiveness of this method. Moreover, a comprehensive evaluation of the CCLN model was conducted using various metrics such as accuracy, confusion matrix, F1-score, and the number of parameters, showing its superiority over other models. Notably, compared to the original signals, the classification accuracy of the AAT-Processed signals increased by 1.14%, validating the effectiveness of this filter method.