Unique Recovery of Motion and Optic Flow via Lie Algebras

A. Peter Blicher, Stephen M. Omohundro

Abstract

This research establishes a rigorous mathematical framework for recovering 3D rigid motion from 2D image sequences using the theory of Lie groups and Lie algebras. By leveraging the differential relationship between temporal changes in image intensity and spatial gradients, the authors demonstrate that 3D motion parameters can be uniquely determined from a minimal set of image points—six for monochrome and three for multi-color images. This approach effectively bypasses the traditional, often error-prone, requirements for explicit point-tracking or feature correspondence.

Introduction

Recovering the 3D motion of objects from 2D visual data is a cornerstone of computer vision. Traditionally, this process involves two computationally expensive and often unstable steps: solving the correspondence problem (matching features across frames) and then estimating motion from these discrete displacements. Furthermore, monochrome systems frequently suffer from the "aperture problem," where motion parallel to an edge remains unobservable at a single point.

This paper addresses these challenges by asking: what is the minimum local information required to uniquely determine the 3D motion of a rigid object? Instead of tracking features over time, the authors propose a differential approach that analyzes the instantaneous change in image values ($df/dt$) relative to spatial gradients.

Methods: Lie Groups and Differential Geometry

The authors model rigid motions in 3D space using the Euclidean group $E(3)$. The instantaneous velocity of a moving object is represented as an element in the associated Lie algebra $e(3)$, which forms a 6-dimensional vector space.

Commutative diagram representing the mathematical framework of rigid motion and image projection.

The image itself is treated as a map from a 2D manifold (the image plane) to an $n$-dimensional color space. The core of the recovery algorithm lies in the equation: $$\frac{d}{dt} f_t(p) = -v \cdot \nabla f_t(p)$$ where the time derivative of image intensity at a point $p$ is the dot product of the optic velocity $v$ and the spatial gradient $\nabla f$. By analyzing the kernel of the projection mapping from the 6D space of possible motions to the available image data, the authors define the conditions for unique motion recovery.

Key Insights: The Power of Color and Genericity

A significant finding of this work is the impact of multi-spectral data. In a 2-color system (e.g., Red and Green), the optic flow vector can be uniquely determined at a single generic point because the gradients of different colors are typically non-parallel. This effectively solves the aperture problem locally.

The authors also introduce the "local spanning lemma," proving that all necessary information for 3D motion recovery is contained within an arbitrarily small image neighborhood, provided sufficient texture variation exists. By treating motion recovery as a linear algebra problem within the 6D $e(3)$ space, they establish that each image point provides a specific number of constraints based on its color dimensionality.

Results: Minimal Point Requirements

The theoretical analysis yields several landmark results:

  1. The 6-Point $df/dt$ Theorem: For monochrome images, the image values and their time derivatives at six generic points are both necessary and sufficient to uniquely specify 3D rigid motion.
  2. The 2-Color Theorem: In images with at least two color dimensions, the optic flow vector is uniquely specified at any generic point.
  3. The 2-Color, 3-Point Corollary: With two or more color dimensions, the requirement for full 3D motion recovery drops from six points to just three.
Geometric interpretation of projecting 3D surface points and local neighborhoods onto a 2D image plane.

These results provide a consistent mathematical bridge between "optic flow" (the temporal paths of points) and the "optic velocity field" (the instantaneous vectors), offering a more rigorous foundation for motion analysis.

Discussion and Limitations

While mathematically robust, the theory relies on several idealizations. It assumes strict object rigidity; deformable objects would necessitate a significantly higher-dimensional Lie algebra. Additionally, the model assumes that intensity changes are purely motion-driven, ignoring photometric effects like specularities or lighting changes. Finally, while the "localness" of the information is proven, practical implementations must contend with sensor noise, likely requiring least-squares estimation over larger image regions rather than infinitesimal neighborhoods.

Summary updated

BibTeX
@inproceedings{blicher1985unique,
  author = {A. Peter Blicher and Stephen M. Omohundro},
  title = {{Unique Recovery of Motion and Optic Flow via Lie Algebras}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure