Vision-based localization and mapping for low-cost MAV applications.

摘要

The past few years have witnessed a marked growing heat over the couple of visual and inertial sensors. Such popularity is gained not only because of its high complementarities between inertial sensors and visual sensors in fundamental motion state estimation, but also due to its advantageous characteristics such as being low cost, low power-consumption, low weight and with small dimension. All these features can be summarized as the capabilities to serve a low-cost, compact, portable and intelligent robot platform. In particular, micro drones, as one of these platforms, are successful representatives taking full advantage of the visual-inertial sensing system on board.;In this thesis, we contribute both a theoretical analysis and a practical implementation of an inertial-assisted visual sensing system. In the algorithm-level, being different from the well-studied visual-inertial sensor fusion, we aim at utilizing inertial sensor readings to simplify the fundamental motion estimation and outlier rejection procedures. In the system-level, we try to work out an embedded solution to the real-time localization and mapping systems for consumer drones, based on our fundamental algorithms. More specifically, the proposed inertial and visual combined system achieves significant improvement in the computational efficiency in fundamental pixel-level algorithms, including the anti-distortion, the stereo rectification, the feature detection and the feature description. On the other hand, based on our 2-point motion estimation algorithm which can simultaneously estimate translational and yaw motion (with given 2D-3D feature correspondences, as well as reliable pitch and roll behaviors from an inertial sensing suite), we thereby design a simple inlier selection scheme which picks the longest sequence of successive correspondences with motion consistency. In the end, several experiments were conducted to demonstrate the feasibility and robustness of our system in complex indoor and outdoor environments.