关键词: 毫米波雷达, 人体姿态估计, 卷积神经网络, 双向长短期记忆网络, 多头注意力机制

Abstract:

Human pose estimation has many applications in human-computer interaction, activity recognition, and health monitoring. Traditional methods based on optical sensors are often limited by lighting conditions and privacy leakage risks, while wearable device-based technologies face issues such as cumbersome usage and discomfort during long-term wear. To address these challenges, a millimeter-wave radar human pose estimation method is proposed that leverages a spatio-temporal fusion framework combining convolutional neural network （CNN）, bidirectional long short-term memory （BiLSTM） network and multi-head attention （MHA）. High-quality point cloud data are generated using self-developed millimeter-wave radar equipment, and a sliding window mechanism is introduced to expand single-frame point clouds into multi-frame time series datas. Spatial features are extracted through CNN, and time-series modeling is performed using BiLSTM, and further optimization of global feature expression through MHA. This spatio-temporal information fusion framework, based on multi-frame point cloud datas, effectively exploits spatio-temporal features, mitigates the radar point cloud sparsity issue, and significantly enhances the accuracy and robustness of pose estimation. Experimental results show that, compared to existing methods, the proposed method successfully localizes all 25 skeletal joints, with average localization errors of 2.69 cm, 2.49 cm, and 2.98 cm along the x, y, and z axes, respectively. This provides a solution for millimeter-wave radar human pose estimation and demonstrates strong practical application potential.

Key words: millimeter-wave radar, human pose estimation, convolutional neural network（CNN）, bidirectional long short-term memory（BiLSTM） network, multi-head attention（MHA）