行人惯性定位新动态:基于神经网络的方法、性能与展望  

Emerging Trends in Pedestrian Inertial Positioning:Neural Network-based Methods,Performance and Prospects

作　　者：李岩[1] 施忠臣 侯燕青 戚煜华 谢良 陈伟 陈洪波[1] 闫野 印二威 LI Yan;SHI Zhong-Chen;HOU Yan-Qing;QI Yu-Hua;XIE Liang;CHEN Wei;CHEN Hong-Bo;YAN Ye;YIN Er-Wei(School of Systems Science and Engineering,Sun Yat-Sen University,Guangzhou 510275;Defense Innovation Institute,Academy of Military Sciences,Beijing 100071)

机构地区：[1]中山大学系统科学与工程学院,广州510275 [2]军事科学院国防科技创新研究院,北京100071

出　　处：《自动化学报》2025年第2期271-286,共16页Acta Automatica Sinica

基　　金：国家自然科学基金(62332019,62076250);国家重点研发计划(2023YFF1203900,2020YFA0713502)资助。

摘　　要：行人惯性定位(Inertial positioning,IP)通过惯性测量单元(Inertial measurement unit,IMU)的测量序列来估计行人的位置,近年来已成为解决室内或卫星信号遮挡环境下行人自主定位的重要手段.然而,传统惯性定位方法在双重积分时易受误差源影响导致漂移问题,一定程度上限制了行人惯性定位在长时间长距离实际运动中的应用.幸运的是,基于神经网络(Neural network,NN)的方法能够仅从IMU历史数据中学习行人的运动模式并修正惯性测量值在积分时引起的漂移.为此,本文对近期基于深度神经网络(Deep neural network,DNN)的行人惯性定位进行全面综述.首先对传统的惯性定位方法进行了简要介绍;其次,按照是否融入领域知识分别介绍了端到端(End-to-end,ETE)的神经惯性定位方法和融合领域知识的神经惯性定位方法的研究动态;然后,概述了行人惯性定位的基准数据集和评价指标,并分析比较了其中一些代表性方法的优势和不足;最后,对该领域需要解决的关键难点问题进行了总结,并探讨基于DNN的行人惯性定位未来所面临的关键挑战与发展趋势,以期为后续的研究提供有益参考.Pedestrian inertial positioning(IP),which estimates a pedestrian's position through measurement sequences from an inertial measurement unit(IMU),has become an important solution for pedestrian autonomous positioning in indoor environments or areas with satellite signal blockages in recent years.However,traditional inertial positioning methods are prone to drift issues during double integration due to the influence of error sources,which to some extent limits the application of pedestrian inertial positioning in long-term,long-distance real-world motion.Fortunately,neural network(NN)-based methods can learn pedestrian motion patterns from historical IMU data and correct the drift caused by inertial measurement values during integration.Therefore,this paper presents a comprehensive review of recent developments in pedestrian inertial positioning based on deep neural network(DNN).First,a brief introduction to traditional inertial positioning methods is provided;Next,the latest research on end-to-end(ETE)neural inertial positioning methods and neural inertial positioning methods incorporating domain knowledge is reviewed;Following that,the benchmark datasets and evaluation metrics for pedestrian inertial positioning are summarized,and the advantages and disadvantages of some representative methods are analyzed and compared;Finally,the key challenges and difficulties that need to be addressed in this field are summarized,and the critical challenges and development trends of pedestrian inertial positioning based on DNN are discussed,aiming to provide useful references for subsequent research.

关 键 词：惯性测量单元 位置跟踪 神经网络 行人航位推算 自主导航 移动设备 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]