Multi-Drone Coordinated Image Sampling for 3D Map Reconstruction through Efficient and Scene-Adaptive Coverage Control

A 3D reconstruction is only as good as the images it is built from — a point seen from one direction reconstructs poorly, seen from several it reconstructs well. Angle-aware coverage control makes that explicit by treating viewing angle as part of the coverage problem. But the existing strategy carries two fundamental problems, and this dissertation is organised around fixing them.

First, it is inefficient on large fields: drones travel long distances to pick off small unsampled patches. Second, and more fundamentally, the measurement granularity has to be fixed in advance — even though how densely you need to sample depends on the scene, which is exactly what you don’t know before flying.

Contribution one — efficiency

The fix for the first problem is to leave the angle-aware objective alone and change what the QP is aiming for: a Voronoi-based coverage law becomes the nominal input to the quadratic program. That keeps the viewpoint diversity angle-aware coverage exists to provide, while giving drones a reason to head for distant unobserved regions — the thing the original formulation had no way to express.

Contribution two — scene-adaptivity

The second problem needs feedback from the reconstruction itself. Images stream into NeuralRecon, which builds an evolving 3D mesh during the flight; changes in that mesh are quantified and used to update the coverage importance index, so regions where the map is still unresolved pull drones back toward them. The objective is therefore shifting underneath the controller as it runs — handled with a QP-based controller that constrains the objective’s decay rate while enforcing collision avoidance and workspace limits.

Validation

Simulated in Unity and ROS 2, and flown on real UAVs both indoors and outdoors. The efficient controller improves monitoring efficiency at large scale; the scene-adaptive method produces more complete and more accurate maps than the equivalent approach without feedback.

The map-feedback line of work is published separately as Coverage-Recon. Full technical detail on the project page: Angle-Aware Coverage Control for 3D Map Reconstruction.