Research On Vision-based Navigation Methods For Aircrafts Using 3D Terrain Reconstruction And Matching

According to the increasing demand of high precision autonomous navigation for Navigation system serves as one of the fundamental components of aircrafts, which supports the flight control and trajectory planning with essential data. Inertial navigation system, global positioning system and terrain(or scene) matching system are the most commonly used state-of-the-art navigation methods. However, the inertial navigation system drifts with time, thus is not suitable for long endurance aircraft positioning and attitude determination; the global satellite positioning system has the problem of anti-interference, which limits its application scope; furthermore, traditional terrain matching technique requires an elevation map built with radar or laser data, as a result, the surveyed area should be relatively large for matching; scene matching based navigation aligns the online map to a pre-built reference map in real-time, which highly depends on the reference map. As a consequence, the problem of autonomous aircraft navigation has become one of the bottlenecks in developing high precision flight platforms.Various aircraft applications, vision-based researches are particularly urgent. However, foreign researchers published very few papers on the area of vision-based aircraft navigation, and domestic research on this topic is still at the preliminary and exploratory stage. This dissertation proposed a novel autonomous aircrafts navigation approach based on 3D terrain reconstruction from image sequence, terrain contour matching and pose estimation. The aim is find feasible and reliable vision-based navigation algorithms in the case of large accumulated errors and GPS jamming, with research focus on image feature matching, real-time parallel dense matching and multi-view relations solver in three-dimensional terrain reconstruction, terrain matching and navigation parameters estimation.The main contents of this dissertation are as follows:1. Proposed a high precision image feature matching algorithm for image sequences with large viewpoint changes.For large view angles, traditional feature matching methods cannot deal with anisotropic deformation, a local patch shape description is proposed to normalize the SIFT descriptor, which can improve the adaption for anisotropic distortion. As known, region based iterative matching is easily trapped into local optimas and has low accuracy. A multi-scale iterative optimization matching method is proposed, which can improve both the efficiency and convergence to the optimal solution with given initial transform parameters.2. Proposed a parallel speedup method for dense image matching.Firstly, according to the characteristics of the imaging sequence, as well as dual-view geometric constraint relationships, homography constraint and epipolar constraint are utilized as the guidance of dense matching, which reduces the search space, as well as improves the computational efficiency and reliability; secondly, since the search procedure under homography and epipolar constraint is suitable for parallel computing, an implementation using the graphics processor(Graphic Process Unit, GPU) has been realized to speed up the dense matching. For the normalized cross correlation matching, the speed is about 200 times faster than its CPU-based implementation; for the affine least square matching method, the speed is about 20 times faster than its CPU-based implementation.3. High precision method of 3D reconstruction of image sequences.Strong maneuverability of the flight platform and repeated textures of the task area often results lower accuracy and less stability. To deal with this problem, we proposed the RANSAC and M-Estimators based fundamental matrix estimation method. Then, as the common transformation model does not adapt to large undulate terrain, we proposed a bundle adjustment approach for image matching and reconstruction, where the gray information has been used as an additional constraint in the adjustment process for the local reconstruction. The reconstructed 3D terrain always suffer from hole points and outliers. To get a more realistic model, a filter and repair algorithm has been proposed for the reconstructed terrain using the geometric constraints between the 3D points. Without attitude parameters of aircrafts, the reconstructed 3D terrain often faces the perspective transformation problem. To deal with this problem, a terrain tilt correction method is proposed based on the topographical distribution of terrain structural.4. Terrain peak matching method based on invariant feature description.Since perspective transformation and similar transformation problems exists between the online reconstructed 3D terrain and the reference map, we proposed a terrain matching algorithm based on an invariant descriptor of peaks. It should be noted that, some region only have the reference map. We proposed a novel scene matching method by resampling the reconstructed 3D terrain from a normal view direction. The corresponding image points and 3D terrain points obtained by terrain or scene matching will suffer from errors inevitably. Moreover, the error may not follow the Gaussian distribution due to terrain relief and undulate. In order to achieve high precision and robust pose estimation for the aircraft, we proposed a weighted function for each matched point pair, and adjust the weighted factor during iterations.