Edge Detection in Optical Flow Fields

William B. Thompson, Kathleen M. Mutch, Valdis BerzinsAAAI 1982

Abstract

This paper addresses the fundamental problem of locating object boundaries using only motion information. By adapting traditional edge detection theories to vector fields, the authors demonstrate a robust way to locate discontinuities in "optical flow"—the pattern of apparent motion in an image sequence—even when data is sparse, noisy, and incomplete.

Introduction

In the early 1980s, researchers recognized optical flow as a powerful source of spatial information. However, they faced a significant challenge: most techniques for calculating flow produced data that was both sparse (only a few matching points between frames) and noisy. Standard derivative-based edge detection methods fail in these conditions because they amplify noise. This paper proposes a method that is "insensitive and adaptive" to these real-world imperfections, allowing for the detection of boundaries regardless of whether motion is caused by the camera or the objects themselves.

Methodology

The authors bridge the gap between scalar edge detection (such as the Marr-Hildreth operator) and vector field analysis through a multi-step process:

  1. Smoothing and Interpolation: To handle sparsity and noise, a Gaussian blurring filter is first applied to fill gaps and suppress high-frequency noise.
  2. Component Decomposition: The vector field is decomposed into its $x$ and $y$ scalar components.
  3. The Laplacian of a Vector Field: A Laplacian operator is applied to each component, approximated by a Difference of Gaussians (DoG)—subtracting a heavily blurred version of the field from a less blurred one.
  4. Zero-Crossing Search: Instead of standard zero-crossings, the authors search for "vector reversals." By analyzing the "deviation" field (the difference between local flow and average flow), edges are identified where deviation vectors point in opposite directions.
Conceptual diagram illustrating vector reversal at a boundary between two moving regions.
  1. Thresholding: To filter false positives, the algorithm enforces thresholds on the angle (must be near 180 degrees) and the magnitude of the combined vectors.

Experimental Results

The method was evaluated using two real-world scenarios involving toy animals (an elephant and a tiger):

  • Case 1: Objects moving at similar speeds but in different directions.
  • Case 2: Objects moving in the same direction but at different speeds.

As shown in the results below, the algorithm successfully traced the boundaries in both cases. In the different-magnitude case (Case 2), which is traditionally harder for many algorithms, the "reversal" logic in the deviation field still allowed for accurate edge identification.

Experimental results showing input frames (a, b) and the resulting motion edge maps (c) for Case 1 (left) and Case 2 (right).

Key Insights and Future Work

A core contribution of this work is the use of the deviation field. By subtracting the local average flow, the algorithm isolates the "signature" of the edge, making it detectable even in complex unconstrained motion scenarios. Furthermore, the authors note that these motion edges carry depth information; the direction of deviation vectors can theoretically indicate which surface is occluding the other.

While successful, the authors acknowledge limitations regarding scale sensitivity—where the success of the operator depends on the size of the Gaussian kernels—and the computational cost of the multi-pass blurring process. Future work would need to address these through multi-resolution approaches and more complex temporal analysis of occlusion cues.

Summary updated

BibTeX
@inproceedings{thompson1982edge,
  author = {William B. Thompson and Kathleen M. Mutch and Valdis Berzins},
  title = {{Edge Detection in Optical Flow Fields}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1982},
}
Figure