DPV-SLAM における AnyLoc を用いた学習ベースのループ検出

This paper proposes DPV-SLAM with AnyLoc loop closure, an enhanced DPV-SLAM that improves loop detection and trajectory accuracy. By replacing the classical BoVW-based loop detection with AnyLoc, a deep learning visual place recognition method, the system remains robust under viewpoint and illumination changes. An adaptive similarity threshold is also introduced to adjust automatically to different environments, removing the need for manual tuning. Experiments on four public datasets show state-of-the-art results on KITTI and TartanAir and clear improvements in accuracy and robustness over the original DPV-SLAM.

In 第31回ロボティクスシンポジア