| 摘要: |
| 现有的人体姿态识别方案大多数是从单一的角度来考察人体的姿态特征，但是仅采用距离像很难体现人体关节的位置信息，仅提取微多普勒特征有时会覆盖掉径向速度不明显的特征。为此，本文首先利用慢时间 距离像和慢时间 微多普勒谱图构建出人体姿态的三维张量数据集，扩展了人体姿态的特征维度，然后采用改进型瓶颈残差模块构成的神经网络提高了人体姿态的识别率。实验结果表明，通过对4名受试者的8种姿态进行训练和测试，该网络对人体姿态的三维张量数据集的识别率可达97.78%，相比于单一特征数据集的识别率提高了4%~7%。 |
| 关键词: 穿墙雷达 人体姿态识别 三维张量数据集 改进型瓶颈残差神经网络 |
| DOI：10.3969/j.issn.1672-2337.2021.01.007 |
| 分类号:TN957.52;TP183 |
| 基金项目:国家自然科学基金(No.61861011,61461012); 广西自然科学基金(No.2017GXNSFAA198050); 广西无线宽带通信与信号处理重点实验室2016主任基金项目（No.GXKL06160106） |
|
| Human Posture Recognition with Multi Dimensional Parameter for Through the Wall Radar |
|
MAO Qiang， JIN Liangnian， LIU Qinghua
|
|
1.School of Information and Communication , Guilin University of Electronic Technology, Guilin 541004,China;2.Key Laboratory of Guangxi Wireless Broadband Communication and Signal Processing, Guilin 541004,China)
|
| Abstract: |
| Most of the existing human posture recognition schemes investigate the posture characteristics of the human from a single perspective. However, it is difficult to reflect the position information of human joints only by range map, and sometimes the nonobvious feature of radial velocity will be covered if we only extract micro Doppler features. To this end, in the paper we firstly construct a three dimensional tensor dataset of human postures by slow time range map and slow time micro Doppler map, which leads to expand the characteristic dimension of human postures. Secondly, the neural network composed of an improved bottleneck residual module is used to improve the recognition rate of human postures. After training and testing eight postures of four human subjects, the experimental results illustrate that the recognition rate of three dimensional tensor dataset of human postures through this network can reach 97.78%, which is 4%~7% higher than the recognition of the single feature dataset. |
| Key words: through-the-wall radar human posture recognition 3D tensor dataset improved bottleneck residual neural network |
