Human Activity Recognition Based on 4D Millimeter-Wave Radar

With the rapid advancement of artificial intelligence, human activity recognition (HAR) has gained increasing attention due to its wide range of applications. To safeguard privacy, the 4D Millimeter-Wave Radar serves as an ideal sensor, offering rich data on coordinates, velocity, and intensity. To improve the spatial information extraction capability of sparse 4D Millimeter-Wave Radar point clouds, we propose a hybrid model combining an improved graph neural network (ATTPoint-GNN) with a Bidirectional LSTM network, designed to process 8-dimensional point cloud data. During the three iterations of the GNN, a Self-Attention mechanism is introduced to focus on the most critical parts of the point cloud motion in the spatial dimension. Additionally, an MLP-based gate mechanism is introduced to update global information, allowing both the vertex and edge information to dynamically incorporate the global information of the GNN. The proposed network structure efficiently classifies 4D human activity point clouds and achieves superior performance on the MMActivity, MMGesture and DGUHA public datasets.