Abstract: Visual-inertial odometry is a widely used localization technology. Visual-inertial odometry technology relies on the assumption of a static environment. In dynamic environments, the robustness and localization accuracy of visual-inertial odometry significantly decrease. Some researchers employ semantic segmentation or object detection methods to identify dynamic objects. However, these approaches face challenges such as the inability to detect undefined dynamic objects, misidentification of static objects, and poor real-time performance. To tackle these issues, we propose a real-time identification method of dynamic feature points to enhance the accuracy of visual-inertial odometry in dynamic environments. This method performs clustering analysis on the velocity vectors of feature points in the image and estimates motion states of feature points by using epipolar matching errors. High-dynamic points are identified and removed, while weight factors are assigned to low-dynamic points. Finally, we evaluate the proposed method on the publicly available datasets. Compared with other visual-inertial odometry algorithms, the preposed approach significantly improves the localization accuracy of visual-inertial odometry in dynamic environments.

Key words: visual-inertial odometry, dynamic objects, localization, feature points, clustering analysis, epipolar matching errors