Visual Odometry

Monocular visual odometry pipeline for real-time camera pose estimation, evaluated on KITTI.

Developed a monocular Visual Odometry pipeline for real-time camera pose estimation using KLT feature tracking, PnP-based localization, and landmark triangulation.

Enhanced the baseline system with:

  • Local Bundle Adjustment for improved trajectory accuracy
  • Ground plane scale correction using SAM3 segmentation
  • Sim(3)-based loop detection and closure

The system was evaluated on the KITTI dataset.

Full pipeline running live on a KITTI sequence (100× speed-up): tracked features on the current frame, landmark counts, relative trajectory error, and the estimated full trajectory.
Final state of the run: the completed KITTI trajectory estimated with KLT tracking and Local Bundle Adjustment.

Code: github.com/nicolejrkim/va25_project