Uncertainty Estimation of Dense Optical-Flow for Robust Visual Navigation

Yonhon Ng, Hongdong Li, Jonghyuk Kim

Abstract

Monocular Simultaneous Localization and Mapping (SLAM) often struggles in environments lacking distinct sparse features or containing repetitive textures. This paper presents a robust framework that utilizes dense optical flow augmented with full 2D uncertainty estimation. By modeling pixel-wise motion as bivariate Gaussian distributions and introducing a Mahalanobis eight-point algorithm for motion estimation, the system achieves superior performance in challenging scenarios like high-speed highways and unstructured forests.

Introduction

Estimating camera motion and scene structure from a single moving camera is a fundamental challenge in robotics. Traditional monocular SLAM systems typically rely on sparse feature points. While efficient, these methods are prone to failure in "featureless" environments—such as highways with uniform barriers or dense forests with complex, repetitive textures—where finding stable, unique corner points is difficult.

Dense optical flow provides a motion vector for every pixel, offering a much richer information source. However, standard dense flow algorithms often assume uniform error across the image. In reality, motion estimation is often anisotropic; for instance, an edge provides strong information perpendicular to its orientation but weak information along it. This work addresses this by quantifying directional uncertainty and integrating it into the motion estimation pipeline.

Methodology

The proposed system builds upon the DCFlow architecture, introducing three primary enhancements to improve robustness and accuracy.

Bivariate Gaussian Uncertainty Estimation

Rather than a scalar confidence value, the system estimates a 2D (bivariate) Gaussian distribution for each pixel's motion. This is achieved by fitting a quadratic surface to the matching cost volume. The resulting information matrix (the inverse of the covariance matrix) captures the directional certainty of the flow estimate.

Figure 6: Estimation of the bivariate Gaussian uncertainty from the matching cost volume. The left image shows the raw cost, while the right shows the fitted distribution.

Mahalanobis Eight-Point Algorithm

The core innovation is the adaptation of the classical eight-point algorithm to account for anisotropic uncertainty. Standard methods minimize the Euclidean distance ($d_E$) between points and epipolar lines. The authors derive a formulation that minimizes the Mahalanobis distance ($d_M$), effectively weighting each correspondence by its estimated 2D uncertainty.

Figure 1: Comparison between Euclidean distance ($d_E$) and Mahalanobis distance ($d_M$) in the presence of anisotropic uncertainty (represented by the gray ellipse).

As shown in Figure 1, the Mahalanobis distance correctly accounts for the shape of the uncertainty distribution, ensuring that noisy components of the motion estimate are down-weighted while reliable components are preserved.

Experimental Evaluation

The framework was evaluated on the KITTI autonomous driving benchmark and a custom Unmanned Aerial Vehicle (UAV) dataset.

KITTI Results

On the KITTI dataset, the method demonstrated improved flow accuracy and, more critically, high robustness. In Sequence 01 (a high-speed highway), traditional sparse methods like VISO2-M failed due to lack of features. The proposed dense approach maintained stable odometry throughout the sequence.

UAV Forest Navigation

Testing in an unstructured forest environment presented challenges such as pure yaw rotations and significant altitude changes. The system maintained a stable trajectory with a scale drift of approximately +5.36%, a notable result for a monocular system without external range sensors.

Conclusion

By explicitly modeling the anisotropic uncertainty of dense optical flow, this research provides a statistically sound method for robust monocular navigation. The Mahalanobis eight-point algorithm allows the system to operate effectively in environments where sparse feature-based methods typically fail. Future work will focus on reducing computational overhead to enable real-time collision avoidance on resource-constrained platforms.

Summary updated

BibTeX
@misc{ng2021uncertainty,
  author = {Yonhon Ng and Hongdong Li and Jonghyuk Kim},
  title = {{Uncertainty Estimation of Dense Optical-Flow for Robust Visual Navigation}},
  year = {2021},
  eprint = {2109.14828},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.14828},
}
Figure