Efficient Visual Inertial Estimation and Control for Micro Aerial Vehicles
微型飛行器的高效慣性視覺運動估計和控制
Student thesis: Doctoral Thesis
Author(s)
Related Research Unit(s)
Detail(s)
| Awarding Institution | |
|---|---|
| Supervisors/Advisors |
|
| Award date | 21 Dec 2021 |
Link(s)
| Permanent Link | https://scholars.cityu.edu.hk/en/theses/theses(90c9f71c-382b-47bf-b746-95e6c2884f3a).html |
|---|---|
| Other link(s) | Links |
Abstract
We have witnessed dramatic progress in the development of Micro Aerial Vehicles (MAVs) in the past decades thanks to their foreseeable applications such as surveillance, transportation and map reconstruction. Autonomous operations of these small drones necessitate accurate and reliable motion estimates. To date, several Visual-Inertial Systems (VINS) have been successfully applied for robot’s localization and mapping tasks. Nevertheless, high computational complexity associated with state-of-the-art VINS prevents them from being deployed on small vehicles with limited power and payload. Thereby, this thesis aims to develop efficient motion and environment estimation as well as control methods for lightweight MAVs with constrained computational resources.
To achieve an efficient motion estimation, we first propose a Kalman-based estimator. The key attribute for efficiency is the adoption of the assumption that only one planar scene is observed by a camera. The method provides the estimates of the orthogonal distance to the plane, plane’s normal, translational velocity, and orientation of the MAVs. Photometric errors are used for the state update to eliminate the feature extraction and tracking process, substantially reducing the computational complexity further. In the second step, the prior reactive navigation regime is extended as a method for Visual-Inertial Odometry (VIO). Apart from the photometric errors, keyframes are introduced. This permits the position and yaw angle of the robot to be estimated. Thirdly, to relax the single-plane assumption employed in the previous two steps, the method is adapted to deal with scenes with multiple planar surfaces. The planes are segmented and tracked across consecutive images through the developed optimization schemes.
For the control task, image-based visual servoing is exploited instead of a position-based approach. The method enables a multirotor robot to track a preset target using only feedback from a monocular camera and a gyroscope. To achieve this, the notion of scaled ratio velocity is considered in place of the absolute translational velocity. This allows the robot to estimate and stabilize its position using only the gyroscope and a camera, without additional sensors to provide metric cues as conventionally needed.
The proposed approaches were validated and assessed with extensive flight experiments. The results indicate that the visual-inertial regimes provide rivaling accuracy at substantial lower computational cost compared to state-of-the-art VINS. Meanwhile, the visual servoing strategy enables the quadrotor to stabilize itself and robustly track a dynamic target. Overall, this thesis offers visual-inertial estimation and control strategies for the micro flying robots with constrained computational resources.
To achieve an efficient motion estimation, we first propose a Kalman-based estimator. The key attribute for efficiency is the adoption of the assumption that only one planar scene is observed by a camera. The method provides the estimates of the orthogonal distance to the plane, plane’s normal, translational velocity, and orientation of the MAVs. Photometric errors are used for the state update to eliminate the feature extraction and tracking process, substantially reducing the computational complexity further. In the second step, the prior reactive navigation regime is extended as a method for Visual-Inertial Odometry (VIO). Apart from the photometric errors, keyframes are introduced. This permits the position and yaw angle of the robot to be estimated. Thirdly, to relax the single-plane assumption employed in the previous two steps, the method is adapted to deal with scenes with multiple planar surfaces. The planes are segmented and tracked across consecutive images through the developed optimization schemes.
For the control task, image-based visual servoing is exploited instead of a position-based approach. The method enables a multirotor robot to track a preset target using only feedback from a monocular camera and a gyroscope. To achieve this, the notion of scaled ratio velocity is considered in place of the absolute translational velocity. This allows the robot to estimate and stabilize its position using only the gyroscope and a camera, without additional sensors to provide metric cues as conventionally needed.
The proposed approaches were validated and assessed with extensive flight experiments. The results indicate that the visual-inertial regimes provide rivaling accuracy at substantial lower computational cost compared to state-of-the-art VINS. Meanwhile, the visual servoing strategy enables the quadrotor to stabilize itself and robustly track a dynamic target. Overall, this thesis offers visual-inertial estimation and control strategies for the micro flying robots with constrained computational resources.
