Robust Keypoint Learning Framework for Visual Odometry in Dynamic Environments

초록

Estimating 6-DoF camera motion in dynamic environments remains challenging because of the influence of moving objects. Traditional feature-based methods and self-supervised approaches struggle with distinguishing dynamic elements from static scenes. To tackle this problem, we propose a robust keypoint detection network designed for pose estimation in dynamic scenes, leveraging cycle consistency and pixel correspondences to guide extracting static keypoints. Our method is validated on highly dynamic sequences from the KITTI multi-object tracking dataset, demonstrating improved performance in dynamic environments.