Research And Implementation Of Fast Visual Navigation Algorithms With Sub-pixel Accuracy

Visual navigation technology is a navigation method that combines inertial navigation information and scene matching to provide flight basis for aircraft.Because of its advantages such as no interaction with the outside and strong anti-interference ability,it is of great significance to the development of UAV navigation technology.However,image matching algorithm,the core technology of visual navigation,has contradiction in speed and precision.The classical SIFT algorithm has sub-pixel accuracy and is easy to implement,but it has a problem that the speed of it is slow.The Key.Net-HardNet algorithm based on deep learning has excellent matching performance,but it only has pixel-level accuracy and no rotation invariance.In order to solve the contradiction between speed and sub-pixel precision of visual navigation system,this thesis carried out in-depth research and innovation from three aspects of pre-storing key-points in base map storage,traditional SIFT algorithm and modern Key.NetHardNet algorithm.The main research results are as follows:(1)Aiming at the problems of long processing time and slow detection speed of the base map in the matching task of real-time image taking by UAV and base map,a new algorithm for pre-storing the base map key-points was proposed.Two strategies were taken to detect keypoints in patches and use KD tree to speed up key-points retrieval.The experimental results based on actual aerial images showed that the matching time reduced by 41.6%(981.50 ms vs.572.88ms)and the number of correct matching points was increased by 4.7%(69.38 vs.72.63)on average.After using the KD tree,the key-points retrieval time was reduced by 75%(4ms vs.1ms).The memory usage of high-resolution base map,which could not be processed by 16 G memory,was stable at about 12% after using the strategy of patch detection,which reduced the hardware requirement for pre-storing key-points of high-resolution base map.(2)Aiming at the problem that SIFT algorithm has sub-pixel positioning accuracy but slow speed,based on the above-mentioned base map key-point pre-storage algorithm,a key-points screening strategy with region of interest and spatial distribution constraints and was proposed,and a complete set of fast visual navigation base on the SIFT was completed.The experimental results showed that the number of key-points of aerial images of different qualities changed gently(about 200)with actual aerial images,and the distribution of key-points in the image was relatively uniform.The visual navigation system successfully matched 106 consecutive UAV images,and the positioning time of a single 500*500 image was less than 600 ms.(3)Aiming at the problem that the key-points detection network Key.Net and the keypoints description network HardNet lack sub-pixel accuracy and rotation invariance,a sub-pixel fitting method based on local gradient fitting and a main direction calculation method based on histogram statistics for pixel-level key-points were proposed.At the same time,affine transformation was performed on the aerial image patches(simulating the flight attitude change of the UAV)to create a dataset,and the Key.Net was trained to further improve the visual navigation performance.The experimental results showed that the sub-pixel localization error of the new Key.Net-HardNet network was less than 0.05 pixels,and the improved network could match rotated images,the matching score after training on the new dataset was increased by 9.2%(43.7% vs.52.9%).