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.
Introduction
Exploring the Martian surface presents a significant robotics challenge due to the substantial round-trip light-time delay, which renders direct teleoperation from Earth impossible. To navigate safely and effectively, a rover must possess high-level "decisional autonomy." This requires the ability to autonomously build a 3D model of its immediate environment to detect obstacles and plan traversable paths.
While active sensors like laser range finders are viable, they often suffer from mechanical complexity or slow acquisition times. This paper explores stereovision—the use of multiple cameras to perceive depth—as a primary on-board sensing modality. Specifically, it addresses the difficulty of matching points between images in unstructured, rocky environments where traditional feature-based stereo methods, which rely on sharp edges, often fail.
Methodology: Trinocular Stereovision and Correlation
The proposed perception system utilizes a trinocular stereovision configuration, employing three CCD cameras arranged in an L-shape. This setup provides redundant information that helps resolve ambiguities in pixel matching.
Trinocular Rectification
Before matching pixels, images undergo a "rectification" process. This transformation ensures that the search for a corresponding point in another image is reduced to a 1D search along a horizontal or vertical line, known as the epipolar line. The authors improved standard rectification methods by selecting specific reference vectors that minimize image distortion, thereby enhancing the performance of the subsequent correlation algorithm.
Correlation-Based Matching
Rather than searching for discrete features like edges, the algorithm employs an area-based correlation approach. A small square window of pixels (typically $5 \times 5$ or $7 \times 7$) is taken from one image and compared against windows in the other images to find the best match. The system evaluates several correlation criteria to determine similarity.
To ensure reliability, a cross-validation check is implemented: a match is performed from Image A to Image B, and then verified by matching from B back to A. If the results are inconsistent, the point is discarded as unreliable, preventing the creation of "hallucinated" obstacles or paths.
Experimental Setup and Datasets
In the absence of actual Martian surface data at the time of research (1992), the team constructed a simulated Martian test-bed. This 16-square-meter outdoor area was filled with sand and rocks of varying sizes (0.4 to 1.2 meters high). To establish a "ground truth" for accuracy assessment, 200 physical marks were manually placed on the terrain. Human operators matched these marks across images to create a gold-standard disparity map for comparison against the automated system's output.
Results and Discussion
The system demonstrated high performance across several metrics:
- Accuracy: The correlation rate reached 95% on the 200 reference points.
- Precision: The mean disparity error remained below one pixel. At a distance of 6 meters, this corresponds to approximately 10 centimeters of depth error, which is sufficient for low-speed autonomous navigation.
- Computational Speed: While a standard 40MHz Sparc processor required 48 seconds for a $256 \times 256$ grid, a specialized architecture using four Digital Signal Processors (DSPs) reduced this to just 6 seconds.
- Robustness: The system handled various lighting conditions effectively, though performance degraded on perfectly smooth sand (due to lack of texture) and in deep shadows.
A key insight was the trade-off in window size: smaller windows are noise-sensitive, while larger windows tend to "blur" rock edges, losing precision. The $5 \times 5$ and $7 \times 7$ sizes were found to be optimal.
Conclusion
The trinocular correlation-based stereovision system provides a robust foundation for planetary rover perception. While it struggles with textureless surfaces and occlusions, the authors suggest that future work should focus on sensor fusion—specifically, integrating stereovision with laser range finders to fill data gaps in shadows or low-texture areas. The project is currently moving toward full integration into the VAP rover backbone for field testing.
Summary updated