基于MIMO雷达成像图序列的切向人体姿态识别方法  

Tangential Human Posture Recognition with Sequential Images Based on MIMO Radar

作　　者：丁传威 刘芷麟 张力 赵恒[1] 周庆[1] 洪弘[1] 朱晓华[1] DING Chuanwei;LIU Zhilin;ZHANG Li;ZHAO Heng;ZHOU Qing;HONG Hong;ZHU Xiaohua(Nanjing University of Science and Technology,Nanjing 210000,China;Shanghai Aerospace Electronic Technology Institute,Shanghai 201109,China)

机构地区：[1]南京理工大学,南京210000 [2]上海航天电子技术研究所,上海201109

出　　处：《雷达学报(中英文)》2025年第1期151-167,共17页Journal of Radars

基　　金：国家自然科学基金(62201259,62301255);中央高校基本科研业务费专项资金(30923011006,30923011026)。

摘　　要：现有的基于雷达传感器的人体动作识别研究主要聚焦于相对雷达径向运动产生的微多普勒特征。当面对非径向,特别是静态姿势或者运动方向与雷达波束中心垂直的切向动作(切向人体姿态)时,传统基于微多普勒的方法无法对径向运动微弱的切向人体姿态进行有效表征,导致识别性能大幅下降。为了解决这一问题,该文提出了一种基于多发多收(MIMO)雷达成像图序列的切向人体姿态识别方法,以高质量成像图序列的形式来表征切向姿态的人体轮廓结构及其动态变化,通过提取图像内的空间特征和图序列间的时序特征,实现对切向人体姿态的准确识别。首先,通过恒虚警检测算法(CFAR)定位人体目标所在距离门,接着,利用慢时滑窗将目标动作划分为帧序列,对每帧数据用傅里叶变换和二维Capon算法估计出切向姿态的距离、俯仰角度和方位角度,得到切向姿态的成像图,将各帧成像图按照时序串联起来,构成切向人体姿态成像图序列;然后,提出了一种改进的多域联合自适应阈值去噪算法,抑制环境杂波,增强人体轮廓和结构特征,改善成像质量;最后,采用了一种基于空时注意力模块的卷积长短期记忆网络模型(ST-ConvLSTM),利用ConvLSTM单元来学习切向人体姿态成像图序列中的多维特征,并结合空时注意力模块来强调成像图内的空间特征和图序列间的时序特征。对比实验的分析结果表明,相比于传统方法,该文所提出的方法在8种典型的切向人体姿态的识别中取得了96.9%的准确率,验证了该方法在切向人体姿态识别上的可行性和优越性。Recent research on radar-based human activity recognition has typically focused on activities that move toward or away from radar in radial directions.Conventional Doppler-based methods can barely describe the true characteristics of nonradial activities,especially static postures or tangential activities,resulting in a considerable decline in recognition performance.To address this issue,a method for recognizing tangential human postures based on sequential images of a Multiple-Input Multiple-Output(MIMO)radar system is proposed.A time sequence of high-quality images is achieved to describe the structure of the human body and corresponding dynamic changes,where spatial and temporal features are extracted to enhance the recognition performance.First,a Constant False Alarm Rate(CFAR)algorithm is applied to locate the human target.A sliding window along the slow time axis is then utilized to divide the received signal into sequential frames.Next,a fast Fourier transform and the 2D Capon algorithm are performed on each frame to estimate range,pitch angle,and azimuth angle information,which are fused to create a tangential posture image.They are connected to form a time sequence of tangential posture images.To improve image quality,a modified joint multidomain adaptive threshold-based denoising algorithm is applied to improve the image quality by suppressing noises and enhancing human body outline and structure.Finally,a Spatio-Temporal-Convolution Long Short Term Memory(ST-ConvLSTM)network is designed to process the sequential images.In particular,the ConvLSTM cell is used to extract continuous image features by combining convolution operation with the LSTM cell.Moreover,spatial and temporal attention modules are utilized to emphasize intraframe and interframe focus for improving recognition performance.Extensive experiments show that our proposed method can achieve an accuracy rate of 96.9%in classifying eight typical tangential human postures,demonstrating its feasibility and superiority in tangential human postur

关 键 词：MIMO雷达 切向人体姿态识别 成像图序列 图像去噪 深度学习 

分 类 号：TN957.52[电子电信—信号与信息处理]