Trajectory Optimization for 3D Scene Reconstruction

Abstract

The work presents an approach to generate a dynamically feasible trajectory for UAVs to perform continuous image acquisition along the trajectory. When given a rough 3D scene proxy, the algorithm generates trajectory and required sensor poses, which yields a high-quality 3D scene reconstruction. We introduce a novel formulation for viewpoint generation, addressing the minimum viewpoint coverage problem in two stages. In the first stage, we solve the linear program with linear constraints that enforce the completeness of 3D reconstruction. In the second stage, we introduce nonlinear integer constraints on the sensor variables, linearize them around the solution of the first stage, and solve a sequence of convex subproblems until convergence. The solution leads to an optimal or near-optimal subset of selected sensors. These viewpoints are then utilized as waypoints to solve a constrained optimal control problem.