Research On The Key Technologies Of Vision Navigation For UAV In Flight

Vision navigation is an important autonomous navigation method. As an effective aidednavigating mode, vision navigation plays an important role on long endurance UAV’snavigation with inertial navigation system during GPS dropouts. Vision navigation is facingthe problem of low robustness, which is influenced by the complicated natural environmentand flying states. Some key technologies of scene matching navigation and motion estimationbased on image sequences are deeply studied in this dissertation. The main contributions areas follows:1. Manual selction of waypoints is a time and labor consumptive job, which is influenced bymany subjective factors. For the above shortcomings, a method of automatic waypointsselection is proposed based on visual saliency analysis to solve these problems. Firstly, thedetection of salient regions with unique structures is implementd by the combination of sparseand low-rank decomposition and sparse coding; then automatic selection of training samplesis implemented by the visual analysis of aerial images, and four feature measures based theedge and cross-correlation surface are extracted from training samples to train SVM model.Aerial images from Google Earth are conducted as experimental samples to design and testthe classifier, and the average classification accuracy is93.33%. A large reference image isobtained via Google Map Downloader for automatic selection of waypoints, and the resultsshow that the method is effective to select waypoints from the reference image.2. For poor robustness of scene matching method based on Hausdorff distance with theinference of rotation and scale, a new scene matching method based on shape context isproposed with weighted Hausdorff distance. Firstly, the key points of the structure areextracted based on the edge continuity detection method to reduce the influence of trivialpoints. Then shape context is used to descript structure information of target edge. The pointmatching cost based on shape context is measured as the weighted coefficient of theHausdorff distance, and a new similar measure is constructed with the constraint of distancefor scene matching. Experiments are conducted on aerial images from Google Earth. Theresults show the method is robust to rotation varying from1o to8oand scale varying from0.9to1.2at the same time. 3. For the problem of larger errors of feature points tracking in image sequence with largermotion, a multi-constraint KLT tracking method is proposed. Firstly, the feature points areselected according to the eigenvalues of optical gradient matrix in the the framework of KLTtracking. A new bi-directional displacement is constructed by integrating both forward andbackward processing. The pyramid model and threshold constrants of point displacements areapplied to optimize the hierarchical estimation of feature points. Image sequences areconstructed via Google Earth images to conduct the tracking experiments. The results showthe proposed algorithm can improve the performance of precise tracking effectively in thecase of large rotation and translation and the performance is better than the methods of P-KLTand TRC-KLT.4. For posterior analysis of feature points matching based on above KLT tracking algorithm, amethod based on topology consistency constraint is proposed to analyze the matching resultsin the image sequence. Firstly, a new distribution graph is constructed, nodes of which arepossible matching points and the edge-weights of which are computed by a new functionbased on weighted shape context within graph point sets and between graph point sets. Then,the points matching in the constraint of topology consistency is realized by an iterativesolution method and the maximum coefficient sum is obtained. At last, the judgement isobtained by the comparision with the coefficient sum of feature points in Chapter4in theabove new graph. The experimental results show that it is effective to determine whether thetracking results are right or not.5. For the problems of large errors caused by the variance of flying environment and states, amotion estimation method based on decomposion of homography matrix is applied toevaluate some factors, such as precision of point detection, image resolution, the quantity anddistribution of feature points, which may affect the estimation of motion parameters. Then themotion estimation method is intergrated with feature tracking in Chapter4and posterioranalysis of homonymous points in Chapter5to construct a navigation method, which isevaluated on images at different resolutions in the visual simulation environment based onVega Prime/MFC. The results provide the experimental basis and introductions for theapplication of motion estimation navigation based on images sequence.6. To meet the requirements of vision aided navigation, an integrated navigationdemonstration and validation system is developed based on waypoints selection method, scene matching method, feature points tracking, and posterior analysis of tracking points.Importantly, some key issues are addressed, such as the acquisition of images with large fieldof view at different time and annotation of geographical information; automatic waypointsjudgement based on INS drift error, size of waypoint and real-time image; relativelocalization in frames based on image mosaicing method in the aviation environment withoutcamera parameters; integrated navigation strategy based on multi-frames consistencyconstraints. The system software is developed on Matlab2012, and these methods proposedin this dissertation are validated with Google Earth images from Xifeng Road in Xi’an.Experimental results show the methods can effectively correct the INS errors and realizehigh-precision navigation.