Visual-Inertial Positioning Method for Dynamic Environment
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.
simultaneous localization and mapping(SLAM)visual-inertial odometrydynamic environment region inpainting
国家科技期刊平台
NSTL
万方数据
