Using Line Correspondence Stereo to Measure Surface Orientation

Lawrence B. Wolff, Terrance E. BoultIJCAI 1989

Abstract

Determining surface orientation is a critical task in robotic vision, yet traditional point-based stereo methods are often sensitive to sensor noise and calibration errors. This paper proposes an alternative approach using line correspondence. By modeling 3D lines as the intersection of projection planes from two cameras, the method provides orientation measurements that are invariant to camera translation and significantly more robust to baseline uncertainty than traditional point-based triangulation techniques.

Introduction

In the late 1980s, robotic vision primarily relied on point correspondence stereo to recover 3D structure. To determine the orientation of a surface, such as a box face, systems would identify multiple points, compute their absolute 3D coordinates, and fit a plane to those points. However, this process is highly sensitive to "baseline errors"—uncertainty in the exact distance between cameras—and pixel-level localization noise. As objects move further from the sensor, depth calculation errors grow quadratically, rendering orientation estimates increasingly unreliable. This research introduces a geometric representation for orientation that is inherently more stable under common sensing errors.

Methodology

The proposed method shifts the focus from discrete points to linear features. A line in 3D space can be represented as the intersection of two "planar sheets of projection." For each camera, a 2D line detected in the image plane, combined with the camera's focal point, defines a unique 3D plane.

Geometry of line correspondence stereo showing planar sheets of projection.

The intersection of these two planes from a stereo pair defines the 3D line. A key advantage of this approach is that the orientation of a line is invariant to translation. Consequently, if the cameras are only translated relative to each other (with no relative rotation), the exact baseline distance is not strictly required to calculate orientation. Surface normals are then derived by taking the cross product of the orientations of at least two coplanar lines, such as the edges of a polygon.

Line representation in the left and right image planes.

Experimental Evaluation

Due to the scarcity of standardized real-world datasets at the time of publication, the authors utilized Monte Carlo simulations to validate their approach. They modeled a triangular object with 15cm edges positioned 1 meter from a stereo rig with a 10cm baseline. Controlled noise was introduced into the simulation, including translational camera uncertainty, rotational errors (±0.5°), and image plane localization noise (±2 pixels). The line-based method was benchmarked against the standard point-based triangulation method proposed by Duda and Hart (1973).

Results and Discussion

The simulations revealed that line-based orientation is far more robust than point-based methods. While point stereo accuracy degrades rapidly with translational uncertainty, the line-based method's error remains relatively stable. This "baseline immunity" stems from the fact that orientation is an angular property, and the intersection of projection planes—also angular constructs—aligns better with the physical sensing process than absolute depth measurement in centimeters.

Furthermore, the line-based approach exhibits graceful degradation at distance. At a range of 2 meters with a 20cm baseline, point-based orientation errors can reach nearly 90 degrees, making them effectively useless. In contrast, the line-based method maintains stability and accuracy. Overall, the line-based method outperformed point stereo by at least 50% for small errors, with the performance gap widening to over 10x as camera uncertainty increased.

Conclusion

While promising, the method faces challenges in stereo matching, as matching lines between two images is more complex than matching points when epipolar constraints are relaxed due to uncertainty. It is also sensitive to the precision of 2D line slopes in the image. Future research directions include fusing line-based orientation with point-based positioning to leverage the strengths of both, and exploring "virtual lines"—imaginary segments connecting corresponded points—to apply these benefits to scenes lacking prominent physical edges.

Summary updated

BibTeX
@inproceedings{wolff1989using,
  author = {Lawrence B. Wolff and Terrance E. Boult},
  title = {{Using Line Correspondence Stereo to Measure Surface Orientation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1989},
}
Figure