Cross-domain human motion recognition

Abstract

One of the most important issues in radar-based pattern recognition is how to efficiently obtain a large amount of reliable labeled data to train the classification model. We train classification models based on simulated samples to classify measured samples across domains to solve the problem of labeled data acquisition in human motion recognition. First, we use the motion capture dataset of Carnegie Mellon University (CMU MOCAP) to simulate the frequency modulated continuous wave (FMCW) radar echo signals of 41 target points of human body. Then generate an amount of simulated labeled data to perform supervised learning to train a Convolutional Neural Network (CNN) classification model. Firstly, a millimeter-wave radar with a two-dimensional antenna array is utilized to transmit FMCW signals and receive the echo signals of human motions. Then data processing is performed on the intermediate frequency (IF) sampling data to gain measured data. Experiment results show that there are similar characteristics in feature spectrograms between those generated by simulation and measurement. Applying the model trained on simulation data to classify the measured data based on a multi-channel CNN with multi-dimensional features method, we achieve test accuracy rate of 94.4%, which proves the feasibility and practicability for cross-domain human motion recognition.