面向毫米波动作识别的视觉辅助信道仿真技术  

Computer Vision-assisted Wireless Channel Simulation for Millimeter Wave Human Motion Recognition

作　　者：任振裕 吉辰卿 余潮 陈万里 王锐 REN Zhenyu;JI Chenqing;YU Chao;CHEN Wanli;WANG Rui(Department of Electronic and Electrical Engineering,Southern University of Science and Technology(SUSTech),Shenzhen 518055,China;Shenzhen Technology University,Shenzhen 518118,China)

机构地区：[1]南方科技大学电子与电气工程系,深圳518055 [2]深圳技术大学,深圳518118

出　　处：《雷达学报(中英文)》2025年第1期90-101,共12页Journal of Radars

基　　金：国家自然科学基金(62171213);高水平专项资金(G030230001,G03034K004)。

摘　　要：该文提出了一种利用计算机视觉技术辅助实现包含运动人体散射特征的毫米波无线信道仿真方法。该方法旨在为毫米波无线人体动作识别场景之下,快速且低成本地生成仿真训练数据集,避免当前实测采集数据集的巨大开销。首先利用基元模型将人体建模为35个相互连接的椭球,并从包含人体动作的视频中提取出人体在进行对应动作时各个椭球的运动数据;其次利用简化的射线追踪方法,针对动作中基元模型的每一帧计算对应的信道响应;最后对信道响应进行多普勒分析,获得对应动作的微多普勒时频谱。上述仿真获得的微多普勒时频谱数据集可以用于训练无线动作识别的深度神经网络。该文针对“步行”“跑步”“跌倒”“坐下”这4种常见的人体动作在60 GHz频段上进行了信道仿真及动作识别的测试。实验结果表明,通过仿真训练的深度神经网络在实际无线动作识别中平均识别准确率可以达到73.0%。此外,借助无标签迁移学习,通过少量无标签实测数据的微调,上述准确率可以进一步提高到93.75%。This study proposes a computer vision-assisted millimeter wave wireless channel simulation method incorporating the scattering characteristics of human motions.The aim is to rapidly and cost-effectively generate a training dataset for wireless human motion recognition,thereby avoiding the laborious and cost-intensive efforts associated with physical measurements.Specifically,the simulation process includes the following steps.First,the human body is modeled as 35 interconnected ellipsoids using a primitive-based model,and motion data of these ellipsoids are extracted from videos of human motion.A simplified ray tracing method is then used to obtain the channel response for each snapshot of the primitive model during the motion process.Finally,Doppler analysis is performed on the channel responses of the snapshots to obtain the Doppler spectrograms.The Doppler spectrograms obtained from the simulation can be used to train deep neural network for real wireless human motion recognition.This study examines the channel simulation and action recognition results for four common human actions(“walking”“running”“falling”and“sitting down”)in the 60 GHz band.Experimental results indicate that the deep neural network trained with the simulated dataset achieves an average recognition accuracy of 73.0%in real-world wireless motion recognition.Furthermore,he recognition accuracy can be increased to 93.75%via unlabeled transfer learning and fine-tuning with a small amount of actual data.

关 键 词：无线信道建模 无线动作识别 无标签迁移学习 毫米波 计算机视觉 

分 类 号：TN957.52[电子电信—信号与信息处理]