Visual Slam for Autonomous Navigation of MAVs
暂无分享,去创建一个
This thesis focuses on developing onboard visual simultaneous localization and mapping (SLAM) systems to enable autonomous navigation of micro aerial vehicles (MAVs), which is still a challenging topic considering the limited payload and computational capability that an MAV normally has. In MAV applications, the visual SLAM systems are required to be very efficient, especially when other visual tasks have to be done in parallel. Furthermore, robustness in pose tracking is highly desired in order to enable safe autonomous navigation of an MAV in three-dimensional (3D) space. These challenges motivate the work in this thesis in the following aspects.
Firstly, the problem of visual pose estimation for MAVs using an artificial landmark is addressed. An artificial neural network (ANN) is used to robustly recognize this visual marker in cluttered environments. Then a computational projective-geometry method is implemented for relative pose computation based on the retrieved geometry information of the visual marker. The presented vision system can be used not only for pose control of MAVs, but also for providing accurate pose estimates to a monocular visual SLAM system serving as an automatic initialization module for both indoor and outdoor environments.
Secondly, autonomous landing on an arbitrarily textured landing site during autonomous navigation of an MAV is achieved. By integrating an efficient local-feature-based object detection algorithm within a monocular visual SLAM system, the MAV is able to search for the landing site autonomously along a predefined path, and land on it once it has been found. Thus, the proposed monocular visual solution enables autonomous navigation of an MAV in parallel with landing site detection. This solution relaxes the assumption made in conventional vision-guided landing systems, which is that the landing site should be located inside the field of view (FOV) of the vision system before initiating the landing task.
The third problem that is addressed in this thesis is multi-camera visual SLAM for robust pose tracking of MAVs. Due to the limited FOV of a single camera, pose tracking using monocular visual SLAM may easily fail when the MAV navigates in unknown environments. Previous work addresses this problem mainly by fusing information from other sensors, like an inertial measurement unit (IMU), to achieve robustness of the whole system, which does not improve the robustness of visual SLAM itself. This thesis investigates solutions for improving the pose tracking robustness of a visual SLAM system by utilizing multiple cameras. A mathematical analysis of how measurements from multiple cameras should be integrated in the optimization of visual SLAM is provided. The resulting theory allows those measurements to be used for both robust pose tracking and map updating of the visual SLAM system. Furthermore, such a multi-camera visual SLAM system is modified to be a robust constant-time visual odometry. By integrating this visual odometry with an efficient back-end which consists…