基于无载波超宽带雷达的小样本人体动作识别

随着雷达硬件平台尺寸越来越小、成本越来越低，室内基于雷达的人体动作识别应用已经成为现实，能够在具有简单架构的低成本设备中实现.无载波超宽带雷达具有极高的分辨力，能够捕获人体细微动作变化并且对室内复杂环境具有很强的抗干扰能力.与基于视频人体动作识别研究相比，超宽带雷达还具有穿透家具、墙体以及保护个人隐私等优点.针对雷达回波信号利用传统时频分析方法实现人体动作识别比较耗时、实时性不好的缺陷，引入机器学习方法对不同类型人体动作进行分类识别.引入机器学习方法用于超宽带雷达人体动作识别最大难点是只有少量可用的超宽带雷达实测数据样本，针对该问题提出基于主成分分析法（PCA）和离散余弦变换（DCT）相结合的人体动作特征提取方法，并利用改进网格搜索算法优化的支持向量机在小样本数据下对人体动作进行识别，最后根据实测数据采取三种不同方案进行仿真实验，结果表明即使在训练数据样本只有5组的条件下，基于PCA和DCT相结合特征提取方法在不同类型人体动作的平均识别率均能达到96%以上.

Few-Shot Learning for Human Motion Recognition Based on Carrier-Free UWB Radar

As radar hardware platforms become smaller and cheaper,indoor radar-based motion recognition applications have become reality and can be implemented in low-cost devices with simple architectures.The carrier-free ultra-wideband (UWB) radar has extremely high resolution,which can capture the slight movement of the human motion and has a strong anti-jamming capability in indoor complex environments.Human motion recognition based on UWB radar compared to video-based also has the advantage of penetrating furniture,walls and protecting personal privacy.Aiming at the defects that the traditional time-frequency analysis method based on radar realizes the human motion recognition is time-consuming and poor real-time performance,the machine learning method is introduced to classify and recognize different types of human motions.The biggest difficulty in introducing machine learning methods for UWB radar human motion recognition is that there are only a few-shot of available UWB radar measured data samples.Therefore,a human motion feature extraction method based on principal component analysis (PCA) and discrete cosine transform (DCT) is proposed.And the support vector machine (SVM) optimized by the improved grid search algorithm is used for human motion recognition under few-shot samples.Finally,simulations experiments are performed based on measured data through three different schemes.Under the condition that there are only 5 groups of training data samples,the average recognition rate of human motion recognition can reach more than 96%.