Indoor 3-D Human Pose Estimation Using Single Millimeter-Wave Radar and Conditional Generative Adversarial Network

In this article, we propose using a single commercial millimeter-wave radar for indoor 3-D human pose estimation. Specifically, we employ chirp signal design and radar data preprocessing methods based on indoor scene optimization to acquire a 5-D radar point cloud of human movements. By extracting multidimensional features of clustered targets from the point cloud and using a machine learning model to identify and eliminate ghost targets caused by indoor multipath, we obtain high-quality, weakly interfered human motion point cloud data. Subsequently, we introduce a conditional generative adversarial network with PointNet as the backbone for reconstructing 3-D human poses.We collect diverse human poses in different indoor scenarios, serving as the training and test sets for the model. Ground truth is collected using the Kinect V2 sensor. Through adversarial training, we enhance the model’s generalization performance for various human poses, achieving high-accuracy 3-D human pose estimation. Finally, the proposed method is validated through experimental results.