| 摘要: |
| 随着隐私保护意识的日益增强与全天候智能感知需求的日益迫切，毫米波雷达凭借其天然的隐私保护特性，已成为人体姿态估计领域的核心替代方案。然而，现有基于雷达谱图的方法通常依赖高分辨率的图像表征，这不仅导致了巨大的数据存储开销，且复杂的网络结构往往伴随着极高的计算复杂度与训练时延。对此，本文提出一种基于轻量型PointNet-LSTM的毫米波雷达点云人体姿态估计方法。首先搭建基于毫米波雷达与深度相机的多模态数据采集系统，构建包含8种典型动作的同步数据集，获取了3D稀疏点云及对应的21个骨骼关键点标签。其次设计了轻量型PointNet-LSTM时空特征提取网络。该网络利用轻量化PointNet模块直接处理原始三维点云，避免了体素化或二维投影带来的量化误差与空间信息损失；同时，引入长短期记忆网络（LSTM）模块捕获帧间的时序依赖关系，学习人体关节的运动规律，以增强姿态估计的连贯性。实验结果表明，所提方法在测试集上的平均关节位置误差仅为 45.4 mm，相较于现有主流方法，本文所提轻量化模型不仅优化了计算效率，还显著提升了姿态重建的准确度。 |
| 关键词: 毫米波雷达 人体姿态估计 点云 神经网络 多模态数据集 |
| DOI： |
| 分类号:TN957 |
| 基金项目:辽宁省科技计划联合计划基金项目（2025110387-JH2/1018） 辽宁省高校基本科研业务费项目（LJ222410143071） |
|
| Lightweight Millimeter-Wave Radar Point Cloud-Based Human Pose Estimation Using PointNet-LSTM |
|
屈乐乐
|
|
|
| Abstract: |
| With the growing public awareness of privacy protection and the increasing demand for all-weather intelligent sensing, millimeter-wave (mmWave) radar has emerged as a promising alternative for human pose estimation due to its inherent privacy-preserving characteristics. However, existing methods based on radar spectrograms typically rely on high-resolution image representations, which not only impose significant data storage overheads but also entail high computational complexity and long training latency due to complex network structures. To address these challenges, this paper proposes a lightweight human pose estimation method based on mmWave radar point clouds using a PointNet-LSTM network. First, a multi-modal data acquisition system integrating mmWave radar and depth cameras is developed to construct a synchronized dataset comprising eight typical human activities, from which 3D sparse point clouds and their corresponding 21 skeletal keypoint labels are obtained. Second, to tackle the non-structural and sparse nature of radar data, a lightweight PointNet-LSTM spatio-temporal feature extraction network is designed. This network utilizes a lightweight PointNet module to process raw 3D point clouds directly, thereby avoiding quantization errors and spatial information loss caused by voxelization or 2D projection. Furthermore, a Long Short-Term Memory (LSTM) module is introduced to capture spatio-temporal dependencies between frames, enabling the model to learn human joint motion patterns and enhance the consistency of pose estimation. Experimental results demonstrate that the proposed method achieves a Mean Per Joint Position Error (MPJPE) of 45.4 mm on the test set. Compared with existing mainstream approaches, our lightweight model not only optimizes computational efficiency but also significantly improves the accuracy of human pose reconstruction. |
| Key words: mmWave radar human pose estimation point cloud neural network multi-modal dataset |