In this paper, we present a low-cost, low-power millimeter-wave (mmWave) skeletal joint localization system. High-quality point cloud data are generated using the self-developed BHYY_MMW6044 59–64 GHz mmWave radar device. A sliding window mechanism is introduced to extend the single-frame point cloud into multi-frame time-series data, enabling the full utilization of temporal information. This is combined with convolutional neural networks (CNNs) for spatial feature extraction and a bidirectional gated recurrent unit (BiGRU) for temporal modeling. The proposed spatio-temporal information fusion framework for multi-frame point cloud data fully exploits spatio-temporal features, effectively alleviates the sparsity issue of radar point clouds, and significantly enhances the accuracy and robustness of pose estimation. Experimental results demonstrate that the proposed system accurately detects 25 skeletal joints, particularly improving the positioning accuracy of fine joints, such as the wrist, thumb, and fingertip, highlighting its potential for widespread application in human–computer interaction, intelligent monitoring, and motion analysis.
In this article, we introduce a novel method for human skeletal joint localization using millimeter-wave (mmWave) radar, effectively overcoming the limitations of vision-based pose estimation methods, which are vulnerable to changes in lighting conditions and pose privacy concerns. The method leverages mmWave radar to generate four-dimensional (4D) time-series point cloud data, which is then projected onto the depth-azimuth and depth-elevation planes. This projection helps mitigate the sparsity inherent in traditional point cloud data and reduces the complexity of the machine learning model required for pose estimation. The input data structure is optimized using a sliding window technique, where consecutive frames are processed by a CNN to extract spatial features. These features are then sorted chronologically and fed into a BiLSTM to capture temporal features, resulting in the accurate localization of 25 skeletal joints. To validate the performance and effectiveness of the proposed method, we created a dataset comprising three body types and 10 distinct actions. The experimental results demonstrate the method’s outstanding human pose estimation capability.