TL;DR
This paper introduces a novel two-stage algorithm for automatically assessing the suitability of landing sites for VTOL aircraft using 3D point cloud data. It moves beyond simple plane fitting to incorporate detailed vehicle geometry and contact simulation, significantly improving landing safety by providing pilots with crucial pre-approach information.
What Problem Does This Paper Address?
Landing VTOL aircraft, especially helicopters, in unimproved or unfamiliar terrain is inherently risky. Current methods often rely on limited visual assessment by the pilot, which is insufficient for detecting subtle but critical hazards like small obstacles or precise slope variations. Prior automated approaches, like plane fitting, only partially capture the complex interaction between the aircraft and the landing surface. This paper tackles the challenge of providing a more comprehensive and reliable automated assessment of landing zone quality, considering both the approach and ground contact phases, to enhance operational safety.
Methods
The proposed algorithm employs a two-stage approach:
- Coarse Evaluation: This initial stage uses a regular terrain grid map to quickly assess potential landing areas. It calculates statistics like slope and roughness for each grid cell. Binary attributes (e.g., sufficient points for plane fit, low point spread, successful plane fit, low residual, and shallow slope) are used to filter out obviously unsuitable areas. This stage is designed to be computationally efficient and incrementally updated.
- Fine Evaluation: Promising sites identified in the coarse stage undergo a more detailed analysis. This involves creating a 2D Delaunay triangulation of the terrain and fitting a 3D model of the helicopter (specifically, its landing skids and underbody) to this triangulation. This allows for simulating the actual contact between the aircraft and the ground, assessing factors like roll, pitch, skid contact area, and potential for underbody or rotor collisions. A goodness score is then calculated based on a linear combination of various factors, including rough evaluation metrics, skid contact, underbody clearance, and distance to obstacles/goals.
Datasets
The authors collected their own 3D point cloud datasets using a helicopter equipped with a Riegl LMS-Q140i-60 ladar scanner, an inertial navigation system (Novatel SPAN INS/GPS), and a documentation camera. Data was collected at various locations around Pittsburgh, PA, USA, covering natural and man-made environments. They performed experiments on 9 sites with a total flight time of 8 hours. No pre-existing public benchmarks were used for evaluation; the results are based on their collected data.
Key Insights and Ideas
- Two-Stage Approach is Crucial: The combination of a fast, coarse evaluation followed by a detailed, fine evaluation is key to balancing computational efficiency with thoroughness. The coarse stage quickly prunes unsuitable areas, reducing the burden on the more computationally intensive fine evaluation.
- Geometric Model is Essential: Moving beyond simple terrain geometry (slope, roughness) to incorporate the specific 3D model of the aircraft (skids, underbody) provides a much more accurate representation of landing feasibility. This allows for detecting issues like tail strikes or insufficient clearance that plane fitting alone would miss.
- Simulating Contact is Powerful: The fine evaluation's ability to simulate the helicopter's interaction with the terrain, by fitting the vehicle model to the triangulated surface, is a significant advancement. This directly addresses the critical ground contact phase of landing.
- Weighting is Subjective but Influential: The final landing site selection relies on a weighted combination of various factors. While the authors manually set these weights, they demonstrate that varying them significantly influences the chosen landing site, highlighting the need for future work in learning optimal weightings.
Key Results
- The algorithm successfully classified landing sites from point cloud maps collected in both urban and natural environments.
- Several potentially good landing sites identified in the coarse evaluation were successfully rejected in the fine evaluation phase due to detailed geometric checks.
- The system can identify landing sites with small slopes and large clearances from obstacles.
- Computation Time:
- Adding points: ~16 ms/100 points (near real-time for their sensor).
- Coarse Evaluation: ~49 ms/1000 cells.
- Fine Evaluation: ~26 ms/cell.
- The fine evaluation is significantly slower, justifying the coarse evaluation's role in limiting its application.
Limitations and Future Work
- Low-Noise Data Assumption: The current method assumes low-noise point cloud data. The authors intend to increase robustness to outliers in future work.
- Water as False Positive: Water surfaces are currently accepted as potentially good landing sites if they provide smooth data. However, they can also result in no measurements, leading to rejection. Future work may involve using other sensor modalities to better distinguish water.
- Vegetation as Roughness: Vegetation is currently classified as roughness, leading to potential false negatives for otherwise good sites. Distinguishing between different types of vegetation and other surface features is a future goal.
- Fixed Landing Gear Model: The fine evaluation currently only models two landing skids. Future work aims to make the algorithm more flexible for different types of landing gear.
- Learning Weights: The weights for the goodness assessment are currently set manually. Future work proposes learning these weights based on human preferences or through self-supervised learning in simulation.
- Scan Matching: To improve registration accuracy and handle potential obstacles smoothed over by current methods, scan matching techniques are suggested for future exploration.
Summary updated