Dynamic Foreground Filtering for Real-Time Stereo Visual Odometry in Urban Traffic Scenes

Abstract

Dynamic foreground objects can degrade feature-based stereo visual odometry by introducing correspondences that violate the static-scene assumption of geometric pose estimation. This paper evaluates three dynamic foreground filtering strategies in a controlled real-time stereo VO pipeline: detection-based semantic masking, optical-flow-based motion consistency filtering, and depth-aware geometric residual filtering. All methods are tested under matched conditions on a real urban stereo sequence recorded with a stereo camera on an experimental autonomous vehicle in Gy˝or, Hungary. The sequence contains 3440 stereo frames, covers approximately 2.02 km, and includes moving vehicles, parked vehicles, pedestrians, and other urban traffic participants. Compared with the unfiltered baseline, all filtering families reduce trajectory error. The best depth-aware configuration decreases raw APE RMSE from 19.68 m to 12.81 m and SE(3)-aligned APE RMSE from 10.12 m to 5.27 m while maintaining real-time operation at the 10 FPS recording rate.