基于时序状态学习模型的行人惯性导航算法

Learning-based pedestrian inertial navigation algorithm with time-series state-space-model

基于深度学习的行人惯性导航方法具有较强的适应性,近年来逐渐成为研究热点。然而,现有方法未充分考虑惯性数据的时序特性,缺乏时序数据的拟合能力。为进一步抑制微惯性导航系统的误差发散,建立了基于时序状态学习模型的行人惯性导航算法。与传统的行人航位推算(PDR)算法不同,此算法不依赖于传统的行人惯性导航框架,而是利用选择性两层双向状态空间模型结构,对特征编码后的隐式惯性特征向量进行时序建模,从而实现位移与不确定性估计。然后,将神经网络估计结果通过扩展卡尔曼滤波进行融合,以进一步降低误差漂移。通过行人导航实验,验证了该方法能够提升定位精度,有效抑制惯性误差发散,实现精准的行人导航。相较于紧耦合可学习惯性里程计(TLIO)方法,绝对轨迹误差和位移漂移率分别降低了32.35%和41.27%。

The pedestrian inertial navigation method based on deep learning has become a major research focus in recent years due to its high adaptability.However,existing approaches do not fully consider the temporal characteristics of the inertial data and lack the ability to fit temporal data.To restrain the error divergence in learning-based micro-inertial navigation systems,a pedestrian inertial navigation algorithm with a learnable time-series state space mode is established.Unlike traditional pedestrian dead reckoning(PDR)algorithms,the proposed algorithm does not rely on the traditional pedestrian inertial navigation framework.Instead,a selective two-layer bi-directional state space model is used to temporally model the implicit inertial feature vectors after feature encoding,and the motion displacement and the uncertainty are estimated.Furthermore,the neural network estimation results are fused by an extended Kalman filter in order to mitigate the error drift.Experiments conducted on wearable pedestrian navigation devices show that the proposed method improves positioning accuracy,effectively suppresses error drift in the inertial system,and achieves reliable pedestrian navigation.Compared to tight learned inertial odometry(TLIO),the absolute trajectory error and displacement drift rate are reduced by 32.35%and 41.27%respectively.