目前，室内场景下诸如定位及追踪、人体活动检测等许多应用对感知的需求越来越大，为支撑高精度感知，高效快速的波达方向(Direction of Arrival，DoA)估计是关键能力之一。然而在具有多径和障碍的室内环境下，传统DoA估计方法存在局限。鉴于机器学习在解决复杂信号处理问题方面呈现的效能优势，对基于深度学习(Deep Learning，DL)的室内 目标DoA估计方法展开研究。设计了专门的卷积神经网络应对未知多径信号带来的不利影响，同时训练了模型从多径分量中分辨提取视距径信号DoA特征的能力。基于此，利用公开的DeepMIMO数据集生成了室内目标DoA估计专用构造数据集来训练模型，提出了基于小样本学习(Few-Shot Learning，FSL)的在线快速学习机制，对提高模型泛化性和鲁棒性展开设计。对室内环境下的DoA估计进行了多种方法仿真对比，验证了所提方法的有效性。
Direction of Arrival Estimation of Indoor Target Based on Deep Learning
Currently,many applications in indoor scenes such as localization and tracking,and human activity detection,have an in-creasing demand for sensing,and accurate Direction of Arrival(DoA)estimation is indispensable for high-accuracy sensing.However,in indoor scenes with multipath and obstruction,traditional DoA estimation methods have limitations.Considering the efficiency advanta-ges of machine learning in solving complex signal processing problems,a Deep Learning(DL)based indoor target DoA estimation meth-od is studied.A specialized convolutional neural network has been designed to cope with the adverse effects of unknown multipath sig-nals,while deliberately training the model's ability to distinguish and extract DoA features of line of sight component from multipath components.Based on this,a dedicated construction dataset for indoor target DoA estimation was generated by using publicly available DeepMIMO dataset.Furthermore,a fast online learning mechanism based on Few-Shot Learning(FSL)was proposed to improve model generalization and robustness.Finally,various methods for DoA estimation in indoor scenes were simulated and compared,verifying the effectiveness of the proposed method.
万方数据
维普
