Abstract
Autonomous navigation in outdoor environments with vegetation is difficult because available sensors make very indirect measurements on quantities of interest such as the supporting ground height and the location of obstacles. We introduce a terrain model that includes spatial constraints on these quantities to exploit structure found in outdoor domains and use available sensor data more effectively. The model consists of a latent variable that establishes a prior that favors vegetation of a similar height, plus multiple Markov random fields that incorporate neighborhood interactions and impose a prior on smooth ground and class continuity. These Markov random fields interact through a hidden semi-Markov model that enforces a prior on the vertical structure of elements in the environment. The system runs in real-time and has been trained and tested using real data from an agricultural setting. Results show that exploiting the 3D structure inherent in outdoor domains significantly improves ground height estimates and obstacle detection accuracy.
Introduction: The Challenge of Unstructured Environments
Navigating a robot through a manicured hallway is a solved problem; navigating a tractor through a field of waist-high tall fescue is a significant challenge. In "unstructured" outdoor environments, sensors like LiDAR rarely hit the actual ground; instead, they return points from the top of the grass or crop canopy. Most existing terrain models treat every patch of ground as an independent unit, leading to noisy estimates and frequent misclassifications of obstacles hidden by vegetation.
The authors argue that we should exploit the inherent structure of the world: the ground is usually smooth, and vegetation in a specific area tends to have a consistent height.
Methods: A Hierarchical Probabilistic Model
The environment is represented as a 3D grid of voxels (15cm³ boxes). To interpret these voxels, the authors propose a hierarchical probabilistic framework that combines vertical and horizontal constraints to "see through" dense vegetation.
- Vertical Structure (HSMM): Each vertical column of voxels is modeled as a Hidden Semi-Markov Model (HSMM). This enforces a logical vertical ordering: Ground $\rightarrow$ Vegetation/Obstacle $\rightarrow$ Free Space. The "Semi-Markov" property allows the model to incorporate a specific prior on the height (duration) of each layer.
- Horizontal Structure (MRFs): Two interacting Markov Random Fields (MRFs) link neighboring columns. One MRF ensures ground height smoothness, while the other maintains class continuity (e.g., if one column is "vegetation," its neighbors likely are too).
- Latent Height Variable: A global latent variable captures the expected vegetation height, allowing the model to share information across the map.
- Inference: The system uses Gibbs sampling to approximate the most likely terrain configuration. Optimized for real-time performance, it runs at 1Hz, sufficient for local navigation.
Experimental Setup
The researchers used a custom-built testbed: an automated John Deere 6410 tractor equipped with dual SICK laser range-finders, high-resolution stereo cameras, an infrared (IR) camera, and high-precision GPS/INS for pose estimation.
Data was collected in real-world agricultural settings, including working farms and undeveloped fields with tall weeds. "True" ground height was established by recording the position of the tractor's wheels as it traversed the terrain.
Key Insights and Results
The core strength of this approach is joint inference. By estimating ground height, vegetation height, and object class simultaneously, the model uses certainty in one dimension to resolve ambiguity in others.
- The "Shed" Test: On a voxel-by-voxel basis, a large white shed might appear as "ground" to a simple sensor model. However, the MRF constraints prevent the ground from jumping 3 meters into the air, forcing the model to correctly classify the structure as an obstacle.
- Accuracy: In challenging tall vegetation where sensors could not penetrate to the soil, the model maintained ground height estimates within ~20cm of the true value.
- Obstacle Detection: The system successfully detected a person hidden in tall weeds, a task where "independent column" models typically fail.
Limitations and Future Work
While effective, the current implementation behaves like a batch process. Future work aims to develop a fully recursive, online algorithm for smoother real-time updates. Additionally, the authors suggest exploring more efficient inference schemes, such as Belief Propagation, to reduce the computational load of Gibbs sampling. Finally, the model currently assumes a simple vertical stack and does not yet handle "hanging" obstacles like tree branches or overhanging cliffs.
Summary updated