环境感知与状态估计是智能网联车关键技术之一.同时定位与建图技术(Simultaneous location and mapping technology,SLAM),旨在同时完成自身的状态估计与环境建模,被广泛应用于智能网联车领域.随着研究的深入,学者们发现通过融合多种传感器,可以实现传感之间的短板互补,提升和加强状态估计的实时性与稳定性.融合视觉和惯性导航仪(Inertial measurement unit,IMU)的实例——视觉惯性里程计(Visual-inertial odometry,VIO),由于具有较高的性价比获得了许多研究人员的青睐.VIO在视觉里程计(Visual odometry,VO)的基础上引入IMU测量,很好地改善了尺度漂移的问题,同时也能极大缓解短期内图像过曝、特征缺失等问题导致的视觉定位失效问题.并且VIO在通过结合冗余传感器提升精度的同时,也通过滑动窗口和状态边缘化等方案保证系统实时性,是兼顾精度和运行效率的典范.细致介绍VIO系统的标准定义与基础模型,并对其关键模块,包括初始化、视觉信息提取与关联、求解与优化、标定,进行详尽的技术梳理与前沿工作回顾,对前沿工作的优点和局限进行详细分析,总结了常用的视觉惯性数据集,并对VIO当前存在的问题和未来发展方向进行了总结和展望.
Survey on Key Techniques for Visual and Inertial Based Odomety
Environmental perception and state estimation is one of the key technologies of intelligent network coupling.Simultaneous location and mapping technology(SLAM),which is widely used in the field of intelligent network connected vehicles,aims to complete its own state estimation and environment modeling at the same time.Scholars in the SLAM field are committed to finding a balance between real-time and accuracy of the algorithm.Visual-inertial odometry(VIO),one of the instances of SLAM schema,is favored by most researchers because of its higher performance and lower price.VIO introduces IMU measurement on the basis of visual odometry(VO),which can not only improve the problem of scale drift,but also greatly alleviate the visual positioning failure caused by image overexposure and feature loss in the short term.As a perceptual measurement with good signal-to-noise ratio,the image can extract high-precision multi view geometric constraints,estimate inertial measurement unit(IMU)bias and noise,and eliminate the cumulative error.Thus,VIO not only improves the accuracy by combining redundant sensors,but also ensure the real-time performance of the system through sliding windows and state marginalization,which is a model taking into account both accuracy and operation efficiency.The standard definition and basic model of VIO system are introduced in detail,and its key modules,including initialization,visual information extraction and correlation,solution and optimization and calibration,are combed in detail and reviewed.The advantages and limitations of frontier work are analyzed in detail,and the commonly used visual inertia data sets are summarized.The existing problems and future development direction of vio are summarized and prospected.
state estimatevisual-inertial odometrysensor fusionsimultaneous localization and mappingvisual-inertial dataset
NETL
NSTL
万方数据
