Local Affine Optical Flow Computation

Hayato Itoh, Shun Inagaki, Ming-Ying Fan, Atsushi Imiya, Kazuhiko Kawamoto, Tomoya SakaiLecture Notes in Computer Science 2014 · DOI · OpenAlex

Abstract

Optical flow estimation is a fundamental task in computer vision, yet traditional local methods like Lucas-Kanade (LK) often struggle with non-constant motion fields. This paper introduces a "local affine" extension to the LK framework. By modeling local displacement as an affine transformation and incorporating $L_1$ regularization via Iteratively Reweighted Least Squares (IRLS), the proposed method achieves robust performance in complex scenarios, particularly the diverging motion fields characteristic of forward-facing vehicle cameras.

Introduction

Estimating the motion of pixels between consecutive frames—optical flow—is critical for tasks ranging from object tracking to autonomous navigation. The classic Lucas-Kanade (LK) algorithm serves as a local "gold standard" due to its computational efficiency. However, LK's core assumption—that motion is constant within a small spatial window—is frequently violated in real-world scenes involving rotation, shearing, or divergence.

While global variational methods (e.g., Horn-Schunck or TV-$L_1$) provide higher accuracy by enforcing global smoothness, they are computationally intensive and often blur motion boundaries. This research explores a middle ground: a local method that enhances expressiveness through affine modeling while maintaining robustness through edge-preserving regularization.

Methodology

The proposed approach upgrades the standard LK framework in two significant ways:

1. Affine Motion Modeling

Instead of a single constant velocity vector $\mathbf{u}$, the displacement is modeled as an affine transform: $$\mathbf{u} = \mathbf{Dx} + \mathbf{d}$$ where $\mathbf{D}$ represents the deformation matrix and $\mathbf{d}$ the translation. This formulation allows the model to capture stretching, rotation, and scaling within the local window.

2. $L_1$ Regularization via IRLS

Standard LK minimizes a least-squares ($L_2$) error, which is highly sensitive to outliers and causes "oversmoothing" at boundaries. The authors introduce an $L_1$-norm regularizer on the affine parameters. Mathematically, minimizing this $L_1$ norm is equivalent to minimizing the Total Variation (TV) of the flow field locally. To handle the non-differentiable nature of the $L_1$ norm, the Iteratively Reweighted Least Squares (IRLS) algorithm is employed.

To accommodate large pixel displacements, the algorithm is integrated into a multi-scale Gaussian Pyramid framework.

Experimental Results

The method was evaluated using the KITTI Vision Benchmark Suite, focusing on front-view driving sequences (Nos. 117, 144, 147, and 181), and the Middlebury Color Chart for qualitative analysis.

Comparison of flow fields: (a) Ground Truth, (b) standard LK 3x3, and (c) the proposed Affine method without regularization ($\lambda=0.0$).

Key Insights

  • Local TV Benefits: By applying $L_1$ regularization locally, the method achieves edge-preserving properties similar to global TV methods without the associated global computational cost.
  • Ego-Motion Suitability: Linearly diverging vector fields, common in automotive contexts, are inherently locally affine. The proposed method is thus highly effective for ADAS applications.
  • Boundary Preservation: The $L_1$ approach significantly reduces the "broken" or noisy vectors seen in standard LK near motion discontinuities.
Qualitative comparison: (d) Ground Truth color map, (e) LK 7x7, and (f) Affine method with $L_1$ regularization ($\lambda=0.5$).

Discussion and Limitations

While the affine model excels in diverging scenes (e.g., sequences 144 and 181), it can be less accurate in purely translational scenes (e.g., sequences 117 and 147) where the additional degrees of freedom may lead to overfitting. Furthermore, the iterative nature of the IRLS solver increases computational overhead compared to the closed-form solution of standard LK. Future work may focus on adaptive parameter selection for $\lambda$ to balance smoothness and data fidelity across diverse motion types.

Summary updated

BibTeX
@inproceedings{itoh2014local,
  author = {Hayato Itoh and Shun Inagaki and Ming-Ying Fan and Atsushi Imiya and Kazuhiko Kawamoto and Tomoya Sakai},
  title = {{Local Affine Optical Flow Computation}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2014},
  doi = {10.1007/978-3-642-53926-8\_19},
}
Figure