Interacting Markov Random Fields for Simultaneous Terrain Modeling and Obstacle Detection.

Carl Wellington, Aaron C. Courville, Anthony StentzRobotics: Science and Systems 2005 · OpenReview

Abstract

Autonomous navigation in unstructured outdoor environments, such as agricultural fields and wilderness areas, presents significant challenges for traditional sensing. Dense vegetation often obscures the load-bearing ground surface, leading to navigation failures. This paper presents a sophisticated probabilistic framework that utilizes Interacting Markov Random Fields (MRFs) to jointly estimate ground height and detect obstacles. By modeling both vertical voxel structures and horizontal spatial dependencies, the system effectively infers the hidden terrain geometry.

Introduction

While autonomous driving on paved roads is a maturing technology, navigating "unstructured" terrain remains a frontier. In environments like cornfields or tall weeds, sensors such as Ladar (laser radar) frequently return signals from the canopy rather than the ground. This leads to two primary issues: the misclassification of sparse vegetation as solid obstacles and the inability to locate the actual driving surface. To address this, we propose a model that leverages spatial "common sense"—the principle that ground is typically smooth and vegetation heights are locally consistent—to "see" through dense cover.

The Probabilistic Framework

The environment is represented as a 3D grid of 15cm³ voxels. To interpret this data, the system integrates three distinct probabilistic components:

  1. Voxel Model: Analyzes individual voxels for density, infrared signatures, and color. It specifically looks for "pass-throughs"—where laser rays penetrate a voxel—to distinguish sparse vegetation from solid objects.
  2. Hidden Semi-Markov Models (HSMM): These model the vertical arrangement of voxels in a column. They enforce logical physical constraints, such as the requirement that vegetation must be supported by ground and that free space exists above the canopy.
  3. Interacting Markov Random Fields (MRFs): These capture horizontal relationships between adjacent columns. A "Smoothness MRF" prevents unrealistic jumps in estimated ground height, while a "Class MRF" ensures that neighboring columns likely belong to the same category (e.g., "corn" or "grass").
The proposed probabilistic framework, integrating (a) voxel-level observations, (b) vertical HSMM structures, and (c) horizontal MRF interactions.

Inference is performed via Gibbs sampling, allowing the interconnected variables to converge on the most likely state of the entire terrain map in real-time.

Experimental Setup

The model was validated using a John Deere 6410 tractor equipped with a comprehensive sensor suite, including roof and bumper-mounted Ladar units, high-resolution stereo cameras, an infrared camera, and high-precision GPS/INS. Testing took place on a working farm and in undeveloped fields with tall weeds. Ground truth was established by tracking the tractor's wheel positions as it traversed the terrain.

Experimental platform: A John Deere 6410 tractor navigating through tall weeds.

Key Results and Insights

The system demonstrated significant improvements over baseline models that treat terrain columns independently. Key findings include:

  • Robust Obstacle Detection: The model successfully identified obstacles, such as a white shed and a person, even when partially obscured by tall weeds. The MRF smoothness constraint was critical here, as it correctly identified vertical jumps in "ground" height as obstacles rather than terrain.
  • Ground Estimation Accuracy: In dense vegetation, the system maintained ground height estimates within approximately 20cm of the true value. By assuming consistent vegetation height, the robot could infer the ground level under a canopy by observing nearby areas where the ground was visible.
  • Real-Time Performance: The system achieved an update rate of 1Hz, which is sufficient for the operational speeds of agricultural machinery.

Conclusion and Future Work

This research highlights the power of combining vertical structural models with horizontal spatial correlations for terrain modeling. Future work will focus on transitioning the current batch-like process into a fully recursive online algorithm and extending the HSMM to handle more complex geometries, such as overhanging tree branches. Additionally, the team is exploring Belief Propagation to further increase inference speed for higher-resolution mapping.

Summary updated

BibTeX
@inproceedings{wellington2005interacting,
  author = {Carl Wellington and Aaron C. Courville and Anthony Stentz},
  title = {{Interacting Markov Random Fields for Simultaneous Terrain Modeling and Obstacle Detection.}},
  booktitle = {Robotics: Science and Systems},
  year = {2005},
  url = {https://openreview.net/forum?id=SCAQMgVC5XY},
}
Figure