| 引用本文: | 孙延鹏，王 爽，屈乐乐，李侠洋. 基于点云数据与微多普勒谱图多模态特征的雷达步态识别[J]. 雷达科学与技术, 2026, 24(3): 289-298.[点击复制] |
| SUN Yanpeng, WANG Shuang, QU Lele, LI Xiayang. Radar Gait Recognition Based on Point Cloud and Micro-Doppler Multimodal Features[J]. Radar Science and Technology, 2026, 24(3): 289-298.[点击复制] |
|
| 摘要: |
| 基于雷达的步态识别作为人体身份识别的一种技术手段具有重要的应用前景，但依赖单模态下的单一特征或同域多特征融合的传统方法不足以支撑鲁棒性的步态识别。随着Mamba网络在计算机视觉领域展现的优越性能，现有的多模态步态识别方法大多缺乏先进的网络架构，识别性能有待提高。此外，例如雷达相机的双传感器模态特征融合类型的步态识别方法在现实中存在设备成本昂贵及隐私泄露问题。为此本文提出一种基于PLVM双流网络的雷达步态识别方法，首先使用毫米波雷达采集人体行走的步态回波数据，经过预处理得到微多普勒时频图以及雷达点云数据，一并送入到PLVM网络进行训练。PLVM双流网络由Pointnet结合双向长短期记忆网络（BiLSTM）分支与具有状态空间模型（SSM）的ViM分支构成，能够有效提取时频图的微多普勒特征和点云的时空特征，将两种模态特征进行互补融合进而完成人体步态识别。实验结果表明相较于传统单模态识别方法，本文方法可有效提升人体步态识别性能，其测试准确率高达98.36%。 |
| 关键词: 步态识别 毫米波雷达 雷达点云 微多普勒特征 Vision Mamba网络 多模态特征融合 |
| DOI：DOI:10.3969/j.issn.1672-2337.2026.03.006 |
| 分类号:TN957.5 |
| 基金项目:国家自然科学基金（61671310）；航空科学基金（2019ZC054004）；辽宁省高校基本科研业务费（LJ222410143071） |
|
| Radar Gait Recognition Based on Point Cloud and Micro-Doppler Multimodal Features |
|
SUN Yanpeng, WANG Shuang, QU Lele, LI Xiayang
|
|
College of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
|
| Abstract: |
| Gait recognition based on radar is a promising technique for human identity recognition. However, traditional methods relying solely on single features in a single modality or on intra-domain multi-feature fusion are insufficient to support robust gait recognition. While Mamba networks have demonstrated superior performance in computer vision, most existing multimodal gait recognition approaches lack advanced network architectures, leaving their recognition capabilities underdeveloped. Furthermore, gait recognition methods based on dual-sensor modality feature fusion—such as those utilizing radar cameras—face practical challenges including high equipment costs and privacy concerns.To enhance this, a novel radar gait recognition approach based on the PLVM dual-stream network is introduced in this paper. Initially, gait echo data from human walking are acquired using millimeter-wave radar. After preprocessing, micro-Doppler spectrograms and radar point cloud data are obtained and fed into the PLVM network for training. The PLVM network comprises a PointNet branch with bidirectional long short-term memory network （BiLSTM） and a Vision Mamba （ViM） branch with state space model （SSM）. It effectively extracts micro-Doppler features from the spectrograms and spatiotemporal features from the point clouds. By complementing and fusing these two modalities, the method achieves human gait recognition. Experimental results demonstrate that this method significantly outperforms traditional single-modal recognition methods, achieving a test accuracy rate of up to 98.36%. |
| Key words: gait recognition millimeter-wave radar radar point cloud micro-doppler features vision mamba network multimodal feature fusion |
