3D Scene Reconstruction with Micro-Aerial Vehicles and Mobile Devices

Scene reconstruction is the process of building an accurate geometric model of one\u27s environment from sensor data. We explore the problem of real-time, large-scale 3D scene reconstruction in indoor environments using small laser range-finders and low-cost RGB-D (color plus depth) cameras. We focus on computationally-constrained platforms such as micro-aerial vehicles (MAVs) and mobile devices. These platforms present a set of fundamental challenges - estimating the state and trajectory of the device as it moves within its environment and utilizing lightweight, dynamic data structures to hold the representation of the reconstructed scene. The system needs to be computationally and memory-efficient, so that it can run in real time, onboard the platform. In this work, we present three scene reconstruction systems. The first system uses a laser range-finder and operates onboard a quadrotor MAV. We address the issues of autonomous control, state estimation, path-planning, and teleoperation. We propose the multi-volume occupancy grid (MVOG) - a novel data structure for building 3D maps from laser data, which provides a compact, probabilistic scene representation. The second system uses an RGB-D camera to recover the 6-DoF trajectory of the platform by aligning sparse features observed in the current RGB-D image against a model of previously seen features. We discuss our work on camera calibration and the depth measurement model. We apply the system onboard an MAV to produce occupancy-based 3D maps, which we utilize for path-planning. Finally, we present our contributions to a scene reconstruction system for mobile devices with built-in depth sensing and motion-tracking capabilities. We demonstrate reconstructing and rendering a global mesh on the fly, using only the mobile device\u27s CPU, in very large (300 square meter) scenes, at a resolutions of 2-3cm. To achieve this, we divide the scene into spatial volumes indexed by a hash map. Each volume contains the truncated signed distance function for that area of space, as well as the mesh segment derived from the distance function. This approach allows us to focus computational and memory resources only in areas of the scene which are currently observed, as well as leverage parallelization techniques for multi-core processing