Monocular Visual Odometry with YOLO-Based Dynamic Object Filtering

• Márcio G. L. Filho ITA
• André O. Françani ITA
• Marcos R. O. A. Maximo ITA

Resumo

This work addresses the challenge of dynamic elements in monocular visual odometry, which traditionally relies on keypoint matching under the assumption of a fully static scene. In real-world environments, however, moving objects violate this assumption and can degrade pose estimation accuracy. To miti-gate this, we propose enhancing a classical feature-based visual odometry pipeline by using the YOLO (You Only Look Once) neural network to detect potentially dynamic objects and remove keypoints located within their bounding boxes. We compare this approach to a baseline method that uses only geometric keypoint matching, using sequences from the KITTI dataset for evaluation. Quantitative evaluation shows an improvement in the accuracy of the estimated trajectory when keypoints on dynamic objects are excluded from the odometry process.

Palavras-chave: YOLO, Accuracy, Filtering, Robot vision systems, Pose estimation, Pipelines, Neural networks, Trajectory, Odometry, Visual odometry, visual odometry, KITTI, monocular camera