Search

Compared to visual sensors that have light dependence and privacy intrusion issues, non-intrusion millimeter-wave (mmw) radars are more suitable for the daily person identification. In a realistic home scenario, there are new challenges that are not taken into account in existing research. This paper attempts to address these issues such as multipath interference, complex walking process and recognition robustness in smart home scenarios, and designs a lightweight multi-branch convolutional neural network with an Inception-Pool module and a Residual-Pool module to learn and classify gait Doppler features. The experimental results in a home living room scenario indicate that the designed mmw radar person identification system can achieve accurate and robust real-time identification performance.

This article presents a robust hand-gesture recognition method via multidimensional feature representation and learning specifically designed for commercial frequency-modulated continuous wave (FMCW) multi-input multi-output (MIMO) millimeter-wave radar. First, the optimal configuration of the radar system parameters for the hand-gesture recognition scenario is investigated and a standard procedure to determine the system configuration is given. Then a moving scattering center model is proposed to represent the 3-D point cloud in the range–Doppler (RD)-angular multidimensional feature space. A scattering point detection and tracking algorithm is presented based on a set of motion constraints in terms of position, velocity, and acceleration. It is derived from the space-time continuity of a nonrigid target. Finally, a lightweight multichannel convolutional neural network (CNN) is designed to learn and classify multidimensional gesture features including radial RD and tangential azimuth–elevation. Extensive experiments are carried out with the developed system and a large data set is obtained to train and test the classifier. The results show that the proposed gesture recognition method can effectively distinguish gestures that are easily confused in the RD domain and achieve robust performances under various conditions.

In this paper, we propose a robust human sleep posture recognition method via multidimensional feature representation and learning based on millimeter-wave (mmw) radar. Firstly, through time-frequency processing of the radar echo signal reflected by the human body, the range spectrum, Doppler spectrum, range-Doppler spectrum, azimuth angle spectrum and elevation angle spectrum of the estimated target are obtained. By setting a fixed frame window length and splicing the above feature spectrums, 5 single-channel 2D radar features are obtained, and combining them in parallel can get a variety of different multi-channel 2D radar feature representations. Finally, a lightweight multi-channel convolutional neural network (CNN) with Inception-Residual module (IRM) is designed to learn and classify multidimensional features. Extensive experiments were carried out using the developed mmw radar system, and a large amount of data was obtained to train and test the classifier. The results show that the proposed sleep posture recognition method can effectively distinguish different sleep postures and achieve better robust performance and generalization compared to other methods.

In this paper, a multi-dimensional feature learning method is proposed to improve the classification accuracy and generalization performance of human behavior recognition on millimeter-wave radar. First, through the process of the radar echo reflected by the human body, we get the spectrums of range-Doppler, Doppler, the azimuth angle and the elevation angle. Then, a fixed frame-length sliding window method is used to capture 6 single-channel image features and 6 three-channel image features that can effectively represent human behaviors. Finally, a lightweight convolutional neural network (CNN) is used to learn and classify multidimensional behavior features. In order to evaluate the effectiveness of the proposed method, a dataset of six classes of human behaviors are collected by 3 people at multiple positions. The experimental results show that, compared with other features, the combined feature of range-time map (RTM), Doppler-time map (DTM) and azimuth-elevation-time map (AETM) has best classification performance for 6 human behaviors. In addition, the data of person A is used to train for a classification model, and the model is used to classify the behavior of people B and C, respectively. The recognition accuracy rates for untrained people B and C are 91.7% and 86.7%, respectively.

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.