Vision-based Landmark Detecton, Position And Orientation Estimation For Mini Unmanned Helicopter

Mini unmanned helicopter has comprehensive applications in military and civilian, and for a 10 years to now, its been as the study object for many institutions and organizations all over the world. Flight control and navigation is the key technology in mini unmanned helicopter. As one type of the advanced navigation, visual navigation holds many advantages such as lower cost, simpler hardware realization and plenty of information contrast to the traditional inertia navigation. Environmental perception, position and orientation of mini unmanned helicopter self-estimation are the two basic requirements of visual navigation. Furthermore, the landmark detection is one of the main means of environmental perception. The work in this dissertation aims at two issues, one is real time landmark detection, and the other is position and orientation of mini unmanned helicopter estimation. The main work of this dissertation is as following:1) In this dissertation, the background and the content of our work are described. Then, a technical overview is presented on landmark detection, position and orientation estimation for mini unmanned helicopter. An analysis and comparison of the convenient methods on this issue are carried out consequently. The advantages and disadvantages of the conventional methods are discussed and the challenges in these issues are described.2) A real-time rotation invariant landmark detection approach which is derived from a face detection approach based on Adaboost algorithm is preseneted. However, unlike the face detection approach which are limited to detecting upright face image, this landmark detection approach which uses the front of series-wound construction classifier as rotation pose estimator not only keep the real-time performance but also detects landmarks at any degree of rotation in the 2D image plane from the air. In the end, results of real flight experiments verify the effectiveness of the approach.3) Without sufficient real training data, the data driven approach which based on Adaboost algorithm cannot solely be utilized in landmark detection. To slove this problem, an approach which uses a boosting algorithm with the prior knowledge for the mini unmanned helicopter landmark detection is proposed. The stage forward stagewise additive model of boosting is analyzed, and the approach how to combine it with the prior knowledge model is presented. The approach is then applied to landmark image detection, where the multi-features are boosted to solve a series of problems, such as rotation, noises affected, etc. Results of real flight experiments demonstrate that for small training examples the boosted learning system using prior knowledge is dramatically better than the one driven by data only.4) A novel monocular vision method for mini unmanned helicopter to estimate its state relative to a known made landmark is proposed. The motion and projection model of the camera onboard of mini unmanned helicopter is established. An original landmark is designed, and several image process algorithms are proposed to extract matching points robustly. In the end, the estimation method applied in a mini unmanned helicopter successfully is introduced.5) A position and orientation of mini unmanned helicopter estimation approach based on the relative information between sequential images is presented. The relationship between sequential images and mini unmanned helicopter movements is first analyzed. Then, based on the analysis, how to use the SIFT features extracted from a few calibrated sequential images to estimate the homography is presented. Furthermore, in order to refine the SIFT features, a feature selection algorithm based on RANSAC is discussed. Finally, the experimental results of real flights verify the position and orientation of mini unmanned helicopter estimation approach effectively.In the end of the dissertation, all the research work is summarized and some further work for deep research is proposed.