Adaptive View Planning for Aerial 3D Reconstruction of Complex Scenes

Cheng Peng, Volkan Isler

Introduction

Reconstructing complex 3D environments using Unmanned Aerial Vehicles (UAVs) presents a significant challenge: balancing limited battery life and onboard processing power with the need for high-fidelity models. Traditional pre-planned flight paths, such as standard zig-zag (lawnmower) patterns, often fail to capture occluded areas or complex geometries, resulting in "holes" or low-resolution regions. While "explore-then-exploit" strategies attempt to solve this by using an initial flight to inform a second targeted pass, they often struggle with highly intricate structures. This paper introduces a novel adaptive view planning method that utilizes "adaptive viewing planes" to efficiently refine 3D reconstructions through iterative data collection.

Methodology: Adaptive Viewing Rectangles

The core of the proposed system is a multi-stage, iterative approach designed to simplify the search for optimal camera positions without sacrificing reconstruction quality.

  1. Initial Exploration: The process begins with a standard zig-zag flight to generate a coarse "proxy" mesh of the scene.
  2. Adaptive Viewing Rectangles (AVR): To avoid the computational expense of searching full 3D space, the algorithm clusters faces of the coarse mesh and fits a set of 2.5D "viewing planes" (rectangles) to these clusters. This approach keeps cameras at optimal distances and angles relative to the geometry while significantly reducing the search space.
  3. Trajectory Planning: The flight path is formulated as a variant of the Traveling Salesperson Problem (TSP). The drone performs back-and-forth sweeping motions on each rectangle, with paths linked via a Minimum Spanning Tree to minimize travel distance.
  4. Iterative Refinement: Following the second flight, the system identifies remaining low-quality regions and plans targeted visits. The authors demonstrate that three rounds of data collection are typically sufficient for high-quality results.
The core methodology involves fitting adaptive viewing rectangles to a coarse proxy mesh to plan an efficient, geometry-aware trajectory.

Crucially, the authors provide a theoretical guarantee that their planned trajectory length is within a constant factor (at most 18x) of the optimal possible path.

Evaluation and Results

The method was evaluated in high-fidelity synthetic environments using Unreal Engine and UnrealCV, testing three challenging scenarios: a natural "GRASS LAND," a complex "OIL Refinery," and a castle-like structure (Scene 3) specifically chosen to test iterative refinement.

The adaptive method significantly outperformed baselines including standard Zig-Zag, Uniform-Grid, and Greedy View Selection (GVS). Key metrics include:

  • Accuracy: Average depth error was reduced by nearly half compared to the zig-zag baseline (104.9 mm vs. 211.5 mm).
  • Completeness: The ratio of pixels with valid depth data doubled from 14.5% (zig-zag) to nearly 30%.
  • Iterative Improvement: Visual analysis confirms that the third visit is essential for filling in complex occlusions that standard two-stage methods miss.
Comparison of reconstruction quality across different methods. The iterative "3rd Visit" significantly improves detail in previously low-quality regions.

Conclusion and Future Work

By utilizing a 2.5D "sweet spot" between simple geometric priors and complex 3D volumetric searches, this adaptive approach provides a practical path for high-quality aerial reconstruction. While the method currently depends on the quality of the initial coarse proxy and has been primarily tested in simulation, it offers a robust framework for autonomous inspection. Future work will focus on integrating real-time SLAM for dynamic positioning and testing the system against real-world environmental factors like wind and variable lighting.

Summary updated

BibTeX
@inproceedings{peng2019adaptive,
  author = {Cheng Peng and Volkan Isler},
  title = {{Adaptive View Planning for Aerial 3D Reconstruction of Complex Scenes}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  url = {https://openreview.net/forum?id=ZD8fcKZdlOd},
}
Figure