面向动态环境的视觉惯性定位方法

Visual-Inertial Positioning Method for Dynamic Environment

针对传统的视觉惯性里程计在动态环境下定位精度低和系统鲁棒性差等问题,提出了面向动态环境的视觉惯性定位方法。首先,利用语义分割提取环境中的语义信息,借助环境先验信息识别出动态物体。同时,采用深度生成网络对动态物体区域进行背景修复,生成只包含静态场景的图像,并将生成的图像用于后续的特征提取和跟踪,以减弱动态物体的影响。后端构建了紧耦合的图优化模型,将视觉数据与IMU数据相互融合,在滑动窗口中以非线性优化的方式估计位姿。实验结果表明,方法可以有效降低动态物体对定位的影响,提高系统的定位精度和鲁棒性。

Aiming at the problems of low positioning accuracy and poor system robustness of traditional visual-inertial odometry in dy-namic environment,a visual-inertial positioning method for dynamic environment is proposed.Firstly,semantic information in the envi-ronment is extracted by semantic segmentation,and dynamic objects are identified with the help of environmental prior information.Meanwhile,the background of dynamic object regions is repaired to generate images containing only static scenes by deep generative network,and the generated images are used for subsequent feature extraction and tracking to attenuate the influence of dynamic objects.The back-end builds a tightly-coupled graph optimization model,which fuses visual data and IMU data,and estimates the pose in a slid-ing window with non-linear optimization.The experimental results show that the proposed method can effectively reduce the influence of dynamic objects on positioning,and improve the positioning accuracy and robustness of the system.