LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry

视觉里程计是基于视觉输入估计移动相机运动的方法。现有的方法主要集中在两个视角点追踪，往往忽略了图像序列中丰富的时间背景，从而忽略了全局运动模式，并且无法对完整轨迹的可靠性进行评估。这些缺点会影响在遮挡、动态物体和低纹理区域等场景下的性能。为了解决这些挑战，我们提出了长期有效的任意点追踪（LEAP）模块。LEAP创新地将视觉、跟踪间和时间线索与经过深思熟虑选择的锚点相结合，用于动态跟踪估计。此外，LEAP的时间概率公式将分布更新集成到可学习的迭代细化模块中，以推断点的不确定性。基于这些特点，我们开发了LEAP-VO，这是一个擅长处理遮挡和动态场景的强大的视觉里程计系统。我们的深思熟虑的集成展示了一种新的做法，即将长期点追踪作为前端。广泛的实验表明，所提出的流水线在各种视觉里程计基准测试中显著优于现有基线。

LEAP-VO: Long-term Effective Any Point Visual Odometry Using Dynamic Anchor

LEAP module combines visual, inter-track, and temporal cues with mindfully selected anchors for dynamic track estimation. Temporal probabilistic formulation integrates distribution updates into a learnable iterative refinement module to reason about point-wise uncertainty.

LEAP-VO is a robust visual odometry system adept at handling occlusions and dynamic scenes. The proposed pipeline significantly outperforms existing baselines across various visual odometry benchmarks. Experiment results are presented and discussed in detail.

Related work in this field includes 'Visual Odometry: Part II - Matching, Robustness, Optimization, and Applications' by Scaramuzza et al., 'Direct Sparse Odometry' by Engel et al., and 'ORB-SLAM: A Versatile and Accurate Monocular SLAM System' by Mur-Artal et al.