微动手势超分辨参数估计及识别技术研究

Fine-grained dynamic gesture recognition technology is a challenging topic in Human-Computer Interaction (HCI) field. It requires the higher estimation precision and correlation of parameters，and higher accuracy of multiple objects tracking, features extraction and recognition of fine-grained dynamic gesture. Studies have demonstrated that the Channel State Information (CSI) and orthogonal multicarrier can be combined to construct a large array of virtual antennas and thus achieve super-resolution estimation of angle and distance of six to eight objects with only three equipped physical antennas. This project aims to study the theory of the super-resolution parameters estimation for fine-grained dynamic gesture based on millimeter wave and dynamic tracking for multiple objects, features extraction and recognition for gesture. First, in order to improve the estimation accuracy of angle, distance, frequency, and the correlation among them, the CSI of millimeter-wave radar and orthogonal multicarrier are employed to build the super-resolution joint estimation model, and the Compression Spectrum (CS) theory is adopted accordingly. Then, in response to interference suppression and similar targets elimination, the theory of nonlinear spectrum sensing and multi-dimensional clustering based on Quantum Harmonic Oscillator Model (QHOM) are studied. Moreover, for the sake of dynamic multi-object tracking, features extraction, and recognition, the dynamic model of Cardinalized Probability Hypothesis Density (CPHD) is built based on random collection. And the fast 5D Convolutional Neural Network (CNN) and adaptive multi-class classifier are then designed. Furthermore, the inequality constrained minimization theory of adaptive scanning beamforming and optimal design of wave form are studied for millimeter-wave radar of fine-grained dynamic gesture. Finally, the demonstration prototype is completed with established systematic theory.

微动手势识别是人机交互的一个挑战性难题，对参数估计精度和关联度，多目标跟踪、特征提取与识别的精准度提出了更高要求。研究发现，利用正交多载波联合信道状态信息可构建大规模虚拟阵列，实现三天线条件下6-8个目标的角度和距离超分辨估计。项目拟开展毫米波微动手势超分辨参数估计，及多目标动态跟踪与特征提取及识别的理论研究。首先，为提高角度-距离-频率参数估计精度及关联度，利用毫米波雷达正交多载波联合信道状态信息构建基于压缩谱理论的超分辨联合估计模型。其次，针对干扰抑制及类目标消除难题，开展非线性空间谱感知、量子谐振子模型结合多维空间聚类的理论研究。再者，针对动态多目标跟踪与特征提取及识别的难点，构建基于随机集的集势概率假设密度动态系统模型，设计高速五维卷积神经网络与自适应多值分类器。最后，开展基于不等式约束极小化的毫米波微动手势雷达自适应扫描波束赋形理论研究及波形设计，形成一套系统化的理论和演示平台。

本项目的总体研究目标是：针对微动手势识别技术难点利用毫米波雷达及其信道状态信息的优势，联合正交多载波技术完成微动手势超分辨参数估计，在此基础上研究电磁杂波及类目标消除技术，进而实现微动手势多目标动态跟踪及特征提取与识别。最后，开展基于不等式约束极小化的毫米波微动手势雷达自适应扫描波束赋形理论研究及波形设计，形成一套系统化的理论和演示平台。研究内容包括了:第一、微动手势超分辨“三域”参数估计理论研究。超分辨距离-角度-频率“三域”参数估计理论是研究微动手势跟踪及特征提取的前提。为提高“三域”参数估计精度，结合微动手势雷达信号的正交多载波、载波相位和信道状态信息，构建基于压缩谱理论的“三域”参数超分辨估计模型，设计平行因子三天线模型提升“三域”参数的关联关系，并提高其估计精度。第二、环境电磁干扰抑制及类目标消除理论研究。环境电磁干扰的抑制是微动手势目标特征提取的前提，类目标消除是微动手势多目标追踪的重要保证，有效提高手势特征提取及手势识别的准确率。第三、微动手势动态多目标跟踪与特征提取及识别。微动手势识别中，在连续的动作序列中实现微动手势目标位置的精确跟踪是提取微动手势时频空特征的前提，微动手势时频空特征的刻画决定了微动手势识别的精度。第四、微动手势识别雷达波形设计及自适应波束赋形研究。微动手势识别需要实现多目标回波信号检测，波形参数会影响目标回波“三域”参数估计效果；为增强手势区域雷达信号作用效果，结合雷达信号的正交多载波和信道状态信息，完成自适应扫描波束赋形设计，提升回波参数提取精度。