mmBehavior: Human Activity Recognition System of millimeter-wave Radar Point Clouds Based on Deep Recurrent Neural Network

Abstract Sensing human activities through wireless technology is the key to human-computer interaction. Relevant studies show that millimeter-wave radar has attracted more and more attention because it can accurately measure distance, velocity, and angle, and extract 3D point cloud images. However, the millimeter-wave point cloud is very sparse. For example, the depth neural network of PointNet is designed to deal with the LIDAR point cloud, which is not effective on the millimeter-wave radar point cloud. Therefore, in this paper, we propose a user activity perception system mmBheavior based on commercial millimeter-wave radar chip. Firstly, we use an improved clustering algorithm P-DBSCAN, which overcomes the problem of DBSCAN parameter sensitivity and realizes adaptive user target detection in the environment of changing number of people. Then we propose a new neural network architecture P-BiLSTMNet, which improves PointNet + + and introduces BiLSTM to learn the time dependence of point cloud. Experiments show that the system can classify all kinds of motion, and the accuracy is more than 93%.