Examination Committee

Prof Ling SHI, ECE/HKUST (Chairperson)
Prof Shaojie SHEN, ECE/HKUST (Thesis Supervisor)
Prof Lu FANG, ECE/HKUST

Abstract

There have been increasing demands for developing micro aerial vehicles with vision-based autonomy in complex environments. In particular, the visual-inertial system (VINS), which consists of only an inertial measurement unit (IMU) and camera(s), forms a great light-weight sensor suite due to its low weight and small footprint. VINS aims at simultaneous localization and mapping (SLAM), which constructs a map of an unknown environment while keeping track of a vehicle’s location. The perception ability equipped with the robots by SLAM technique plays an essential role during autonomous flight.
 
In this work, we study two parts of the SLAM problem: localization and mapping.
 
For localization, we develop efficient, high-accuracy VINS using probabilistic graph model. Towards plug-and-play and highly customizable VINS, we extend our system to address two challenges: the initialization problem and the calibration problem. We propose a methodology that is able to initialize velocity, gravity, visual scale, and camera-IMU extrinsic calibration on-the-fly. Our approach operates in natural environments and does not use any artificial markers. It also does not require any prior knowledge about the mechanical configuration of the system. The proposed approach also allows generalizing the monocular VINS to multi-camera VINS, which significantly boosts the system’s accuracy and robustness. We made comprehensive experiments in large-scale indoor and outdoor environments to demonstrate the performance of our system.
 
Mapping can be treated as a dual problem of localization, as localization makes mapping feasible and mapping is able to reduce the drift of localization. Building on top of the localization system, we develop a scalable monocular mapping system to construct and update the surrounding environments. By utilizing device’s GPU the system achieves dense 3D reconstruction at frame rates. Experiments in both indoor and outdoor environments prove the reconstruction quality is comparable to systems that need depth sensor or stereo cameras.