A Novel Solution to Vision-Based Next-Best-View Selection for 3D Reconstruction with Limited A Priori InformationHenner, Coleman NBVSLAMYOLOAIUAVNext-Best-ViewPhotogrammetryLiDARYou Only Look OnceComputer VisionRemote SensingUAV AutonomyVision-Based SystemsPath PlanningAutonomyPoint CloudTrajectory OptimizationAerospace EngineeringB SThis work presents a novel approach to the Next-Best-View (NBV) problem for UAVs solely equipped with a monocular RGB camera and with limited a priori knowledge of their target. The vehicle first performs a predefined search routine to locate and obtain several initial views of a target using the You Only Look Once (YOLO) artificial intelligence object classification framework. An implementation of Monocular Visual Simultaneous Localization and Mapping (SLAM) is used to estimate corresponding camera poses and generate a sparse point cloud of the scene. A 3D bounding volume, wherein the target is expected to be located, is defined by projecting YOLO detection bounding boxes into the point cloud space. Subsequent viewpoint selection is informed by optimizing an information gain heuristic at a set of dynamically generated candidate viewpoints surrounding the bounding volume. YOLO and SLAM are run simultaneously on newly acquired images to incrementally update the point cloud and bounding volume. As the vehicle obtains more information about its surroundings, the optimal trajectory is recalculated for the NBV. This approach is validated in a virtual environment simulating the optics and dynamics of a UAV. Coverage of the target converged to 85% within 35 frames, and the estimated target surface area reached a steady value by 25 frames.Simon W Miller, Thesis SupervisorMark David Maughmer, Thesis Honors Advisoropen_access2025-04-09T21:51:27Zhttps://honors.libraries.psu.edu/catalog/9706ckh5586