On-board visual control algorithms for unmanned aerial vehicles

Abstract

Abstract: The aim of this thesis is to exploit the vast amount of information given by a visual sensor in order to design visually guided Unmanned Aerial Vehicles. This work, goes beyond the traditional use of visual systems on UAV, of be a simple ¿eye in the sky¿, by demonstrating that a visual sensor provides trusted and reliable information useful for the design of different control system and applications and with a capacity, precision and performance similar to the one demanded for traditional control sensors. This thesis presents at first how the state of the art visual tracking and image processing algorithms are robustly and efficiently implemented on-board UAV for estimation of different projective transformations. These transformations (homographies included), are then used for aerial images enhancement, presenting a real time video stabilization and mosaic building system. The properties of the projective transformation are then employed to build a novel 3D pose measure system, that allows to obtain the camera-aircraft metric position and attitude w.r.t. a ground reference plane. The metric and attitude position estimated by the homography decomposition, are used to generated a high level decoupled pose control system that correctly drives the non-holonomic behavior of a rotary wing UAV and the common challenges presented on aerial images like higher vibrations, illumination changes and visual references occlusions and large displacements, among others. The presented contribution is used for UAV control on situations like autonomous landing and accurate hover positioning. High Level decoupled control strategies are also used to develop an Image Based Visual Servoing architecture for a novel aerial moving object following system, in which the analysis of the interaction between the image references and the system degrees of freedom DOF are used to reduce the effects of the aircraft dynamics on the visual control design and tuning. This thesis also presents how Omnidirectional systems are used to estimate the aircraft flying attitude and to make a visual compass. The proposed real time method uses the properties of the unify theory for central catadioptric cameras in order to build an attitude sensor based on the skyline segmentation and the motion register between sequential omnidirectional images. Finally, the properties of a decoupled visual control and an omnidirectional system are integrated in order to generate a novel collision avoidance system. This contribution on the field of See&Avoid area, uses a decoupled Image base visual architecture designed on the unitary sphere manifold, allowing to directly include the properties and advantages of a large field of view FOV given by omnidirectional sensor in the design of a visual control systems for UAV. All the proposed method are tested and validated on real aerial test using rotary wing UAV. The contributions presented on this thesis can be easily extended to other kind of UAV as well as general mobile robotics. This thesis demonstrates that computer vision onboard UAV is an unexploited field, that has proved to be a versatile, low cost and effective sensor systems suitable for many industrial applications as well as a principal component for UAV safety operation. This thesis opens the doors to the development of novel system for industrial and civilian markets in which there is necessity of use visual data to increase aircraft autonomy and capabilities guaranteeing a safety operational levels.