交互式多模型算法改进视觉惯性定位技术  

Improved visual-inertial positioning technology by interactive multiple model algorithm

作　　者：陈锐锋 赵立业[1] CHEN Ruifeng;ZHAO Liye(School of Instrument Science&Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学仪器科学与工程学院,南京210096

出　　处：《导航定位学报》2024年第5期80-88,共9页Journal of Navigation and Positioning

摘　　要：针对目前使用优化方法构建的视觉惯性数据融合模型尚未充分考虑实际复杂性,导致无法准确模拟复杂的实际状态的问题,提出一种交互式多模型(IMM)算法改进视觉惯性定位技术:指出相较于传统滤波方法,优化的视觉惯性融合定位技术具有更高的估计精度,并且在计算效率上能够与滤波方法媲美,这得益于雅可比矩阵和海塞矩阵的稀疏性特征;然后给出将交互式多模型算法与优化算法相结合的方法,利用IMM能够有效模拟单个目标在多种情景下的状态的特点,通过模型概率实时分配惯性数据和视觉数据的置信度,改进优化算法,从而提高位姿估计精度。实验结果表明,相比基于旋转不变特征点的实时定位构图系统第三代(ORB-SLAM 3)算法,该方法对定位精度的均方根(RMS)误差性能可提升17%。Aiming at the problem that visual-inertial data fusion models built by using optimization methods have not fully considered the actual complexity,resulting in the inability to accurately simulate complex real-world states,the paper proposed an improved visual-inertial positioning technology by interactive multiple model algorithm:it was pointed out that compared to traditional filtering methods,due to the sparsity characteristics of Jacobian and Hessian matrices,optimization techniques could offer higher estimation accuracy and compete with filtering methods in computational efficiency;then,the interactive multiple model(IMM)algorithm was combined with optimization algorithms,utilizing the capability of IMM to effectively simulate the state of a single target under various scenarios,and dynamically assigning the credibility of inertial and visual data by model probabilities,to improve the optimization algorithm and enhance the pose estimation accuracy.Experimental results showed that compared to the oriented brief-simultaneous localization and mapping 3(ORB-SLAM 3)algorithm,the proposed method could improve the root mean square(RMS)error performance of localization accuracy by 17%.

关 键 词：交互式多模型算法 组合导航 优化 融合定位 视觉定位 惯导 

分 类 号：P228[天文地球—大地测量学与测量工程]