Autonomous planetary rover (VAP): on-board perception system concept and stereovision by correlation approach

L. Boissier, Bernard Hotz, Catherine Proy, Olivier D. Faugeras, Pascal Fua

TL;DR

This paper outlines the perception architecture for the French "Autonomous Planetary Rover" (V.A.P.) program, designed for Mars exploration. Its core contribution is a robust trinocular stereovision system that uses a correlation-based approach to generate 3D maps of rugged terrain, achieving high accuracy and near-real-time performance on 1990s-era hardware.

What Problem Does This Paper Address?

Exploring Mars presents a massive robotics challenge: the round-trip light-time delay makes direct teleoperation impossible. A rover must possess a high degree of "decisional autonomy" to navigate safely.

To do this, the rover needs to build a 3D model of its environment to detect obstacles and plan paths. While laser range finders are an option, they are often slow or mechanically complex. This paper explores stereovision—using multiple cameras to perceive depth—as a primary on-board sensing modality. Specifically, it tackles the difficulty of matching points between images in unstructured, rocky environments where traditional "feature-based" stereo (which looks for sharp edges) often fails.

Methods

The authors propose a trinocular stereovision system (using three CCD cameras in an L-configuration) and a correlation-based matching algorithm.

  1. Trinocular Rectification: Before matching pixels, the images must be "rectified" so that searching for a corresponding point in another image is reduced to a 1D search along a horizontal or vertical line (the epipolar line). The authors improved upon standard methods by choosing specific reference vectors that minimize image distortion, ensuring the correlation algorithm performs better.
  2. Correlation Approach: Instead of looking for edges, the algorithm takes a small square window of pixels from one image and searches for the most similar window in the other images.
  3. Validation: To prevent "false matches," they use a cross-validation check: they match from Image A to Image B, then from B back to A. If the results don't agree, the point is discarded as unreliable (marked as black/unmatched).
  4. Hardware Acceleration: Recognizing that these calculations are intensive, they tested the algorithm on a Sun 4 workstation and a specialized architecture featuring four Digital Signal Processors (DSPs).

Datasets

The researchers didn't have access to actual Mars data in 1992, so they built a simulated Martian test-bed:

  • Physical Setup: A 16-square-meter outdoor area filled with sand and rocks ranging from 0.4 to 1.2 meters high.
  • Ground Truth: To measure accuracy, they manually fixed 200 physical marks on the terrain. Humans matched these marks across images to create a "gold standard" disparity map to compare against the computer's results.

Key Insights and Ideas

  • Distortion Matters: The math for image rectification is well-known, but the authors realized that how you project the images into a common plane affects the correlation. By aligning the rectification vectors with the camera baselines, they minimized "stretching" the images, which significantly improved matching accuracy.
  • The Window Size Trade-off: They found that $5 \times 5$ and $7 \times 7$ pixel windows provided the best balance. Smaller windows are too sensitive to noise, while larger windows "blur" the edges of rocks and lose precision.
  • Density vs. Reliability: By using the cross-validation check, the system prioritizes being "correct" over being "complete." It’s better for a rover to see a "black hole" in its data (indicating uncertainty) than to hallucinate a flat path where a rock actually exists.

Key Results

  • Accuracy: On the 200 reference points, the system achieved a 95% correlation rate.
  • Precision: The mean disparity error was less than one pixel. In their setup (6 meters from the robot), a one-pixel error translated to roughly 10 centimeters of depth error—sufficient for slow-speed navigation.
  • Speed: On a 40MHz Sparc processor, a $256 \times 256$ grid took 48 seconds. However, using four DSP processors, they brought this down to 6 seconds, approaching the speeds necessary for practical field use.
  • Robustness: The system successfully handled various lighting conditions, though it struggled with perfectly smooth sand (which lacks texture for correlation) and deep shadows.

Limitations and Future Work

  • Textureless Surfaces: As noted, the correlation method fails on smooth sand because there are no unique patterns to match.
  • Occlusions: While the third camera helps, "blind spots" behind large rocks remain a challenge.
  • Sensor Fusion: The authors suggest that the next step is "cooperation" between stereovision and laser range finders—using the laser to fill in the gaps where the cameras struggle (like in low-texture areas or shadows).
  • Integration: The paper concludes by noting the project is moving toward integrating these perception modules into the full "rover backbone" for real-world testing.

Summary updated

Abstract

The authors present some of the work currently being conducted at CEA, CNRS, INRIA, ONERA, and CNES in the framework of the Autonomous Planetary Rover program (VAP). A system concept approach for the VAP application is first presented. Attention is then given to the VAP perception system: some results obtained with stereovision by a correlation algorithm on outdoor simulated Mars terrain scenes are shown.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>
BibTeX
@inproceedings{boissier1992autonomous,
  author = {L. Boissier and Bernard Hotz and Catherine Proy and Olivier D. Faugeras and Pascal Fua},
  title = {{Autonomous planetary rover (VAP): on-board perception system concept and stereovision by correlation approach}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {1992},
  doi = {10.1109/robot.1992.220265},
}

Loading related content...

Figure