Research On Several Problems Of Vision-based Navigation And Control For Aerial Robotics

Aerial Robotics, also known as Unmanned Aerial Vehicles (UAV), have been an active area of research at home and abroad in these decades, because they can be widely used in many civil and military applications. The implementaion of vision-based navigation and control, which can manage tasks of obstacle avoidance, search and rescue, aerial surveillance, is one of the prerequisites for UAV to be widely applied in reality. However, there are many difficulties when processing the image sequences captured from UAV, including maneuver, illumination changes, and viewpoint changes of the cameras, etc.In this dissertation miniature unmanned helicopter(MUH) is selected as the experimental platform, and some researches have been done to solve the problems in vision-based navigation and control for aerial robotics, including object tracking in search and rescue scenarios, landmark detection, relative pose estimation and navigation signal computation in autonomous landing scenarios, etc. The main work and contributions of the dissertation are listed as follows:1) The methods of vision-based navigation and control for robotics are systematically reviewed. The recent advances in vision-based navigation and control for aerial robotics are introduced, and some typical applications are listed. Serveral problems in vision-based navigation and control for aerial robotics are analyzed, and the work of the disseration is presented.2) A tracking algorithm of ensemble tracking embedded particle filter (ETEPF) is proposed. The performances of three existing algorithms, which use kernel histogram as observations, are first compared. It shows that the combination of mean shift and particle filter can improve tracking performance. Ensemble tracking (ET) is then embedded into a stochastic tracking framework to solve the difficulties of abrupt motion and illumination changes, etc. It chooses likelihood mean and kernel histogram as observations, and updates motion models online to depict the evolution of particle's states. The experimental results demonstrate that, ETEPF reduces tracking errors, can recapture the object quickly when partial or full occlusion occurs, and achieves robust object tracking in complex environments.3) To overcome high noise and motion blur in landmark sequences captured from MUH, an algorithm of landmark detection and tracking based on layered particle filter is proposed considering the temporal and spatial consistency of landmarks in successive frames. The local contrast descriptor is proposed as the feature vector for each pixel to remove the influences of uniform illumination changes and reduce the computational cost of feature extraction, and ensemble method is used to discriminate patterns between landmarks and backgrounds. The layered particle filter considers the hierarchy of the used features, and improves real-time performance. The experiments using real sequences demonstrate the effectiveness of the proposed algorithm.4) To solve the image registration problem in relative pose estimation based on homography, the matching and real-time performances of two local features, including SIFT and SURF, are compared. The experimental results show that SURF has the same advantages of quantity and distinctiveness as SIFT, but imporves real-time performance of feature extraction and matching, and can meet requirements in visual navigation for aerial robotics. The effectivenss of the pose estimation algorithm based on homography is also proved through experiments,5) A hybrid visual servoing scheme using image moments (MHVS) is proposed to solve the difficulties of changes of matched local features during the servoing process. This scheme estimates homography through point correspondence first, and then computes the pixel coordinates of unmatched points by back projection. The moments of local features and the relative poses through homography decomposition are used to control translational and rotational degree of freedom of the robotics, respectively. The jacobian matrix of the corresponding features is not singular, and the control signal does not need the computation of matrix inversion, so it is especially suitable for actual applications. The stability and robustness to camera calibration errors are analyzed theorectially. The simulation results show that MHVS improves servoing performance, and is applicable for local features.6) A switching control scheme is presented to solve the difficulties of object's leaving the camera's view during the servoing process, in which two hybrid visual servoing schemes, including KD and MHVS, are switched according to the object's position in image plane to manage servoing tasks. The simulation results show that the proposed scheme improves the robustness of the servoing systems to some extent.The current work is summarized and some further researches are presented at the end of thedissertation.