Examination Committee

Prof Ming LIU, ECE/HKUST (Chairperson)
Prof Chi Ying TSUI, ECE/HKUST (Thesis Supervisor)
Prof Ling SHI, ECE/HKUST
Prof Kam Tim WOO, ECE/HKUST

Abstract

In recent years, autonomous aerial robots are becoming popular for research and for commercial and industrial. Many applications, for aerial robots, such as delivery, infrastructure inspection, and surveillance, involves operations at high altitude. Nowadays most of the state estimation problems at high altitude are solved by a fusion of Inertial Measurement Unit (IMU) and GPS. However, high altitude does not guarantee good GPS reception. In fact, for operations in urban areas that involve flying between high-rise buildings or in the middle of deep canyons, GPS is often blocked due to obstructed sky view. There is certainly a high demand in developing state estimation solutions that work at GPS-denied/downgraded high altitude environments. Due to the lack of direct distance measurements, monocular visual-inertial solutions become attractive.
 
The problem of visual perception can be divided into two parts: localization and depth reconstruction. For localization, there exists many visual inertial estimators which achieve accuracy and robustness. However, these estimators suffer from initialization under poor numerical conditioning or even degeneration at high altitude, due to difficulties in retrieving observations of visual features with sufficient parallax, and the excessive period of inertial measurement integration. A spline-based high altitude estimator initialization method for monocular visual-inertial navigation system (VINS) is proposed in this work to tackle the aforementioned initialization issues. 
 
For monocular depth reconstruction, existing methods usually assume that the localizing results are accurate. But actually the estimators are not guaranteed to always give out accurate results at high altitude, so it is reasonable to compensate for the inaccuracy at the depth sensing stage. Different from 1-Dimension cost evaluation, we make use of the recent results from optical flow community as the front-end for monocular depth reconstruction. Poses are firstly refined through structureless bundle adjustment, then the depths are estimated from the optical flow results and the refined poses of several consecutive frames.