Research On Registration Algorighms Of Scene Matching Aided Navigation System Under Multi-source Information Fusion Frame

Scene matching aided navigation system(SMANS) is used to reduce the cumulative error of Inertial Navigation System(INS). It is a kind of active navigation mode which is an efficient backup for Global Positioning System(GPS) and can cut back the signal exchanges between the jet and the outside to keep its stealth. As an important application of multi-source information fusion technology, the accuracy of SMANS depends on the precision of image registration directly, so we study the registration methods of scene images which are from different sensors and with a certain degree of distortion.The main contents and innovations are as follows.(1) A Multi-scale Auto-convolution(MSA) registration method is proposed based on a kind of affine invariant feature. In this method, a simplified Euclidean distance is used as similarity measure to accelerate the matching progress. In addition, a three-level pyramid hierarchical searching strategy is used so that the computational complexity is reduced to one-eighth of the original. Compared with the Hausdorff matching method, the MSA method can efficiently resist the distortion of sensing image and acquire more accurate locating result for SAR, CCD and IR image.(2) Scale Invariant Feature Transform(SIFT) is proved to be rotation and scale invariant but it is not stable for shear transformation. Based on the idea of feature fusion we combine SIFT and MSA together to form a more stable feature for affine transformation. After the stable SIFT feature point is acquired, we calculate the MSA feature of the circular area with center at the key-point and radius proportional to magnitude of main direction gradient in order to ensure the rotation invariance. All ?SIFT+MSA? features of sensing image and target image are calculated and ratio threshold method is chosen to find coarse matches. Then we apply the Random Sample Consensus(RANSAC) method to get rid of the mismatched point pairs to get more accurate homography matrix. Compared with the SIFT registration method, the SIFT+MSA feature has higher feature extraction efficiency and lower RMSE value after registration.(3) In order to quantitatively measure the effectiveness and accuracy of the proposed algorithm, we build the evaluation system of registration method from three aspects, which are feature extraction efficiency, accuracy of registration result and enlarged details of result image.(4) In order to assess the feasibility of implementing the algorithm on engineering practice, explore the fusion solution of scene matching aided navigation and inertial navigation data link, and intuitively demonstrate the INS cumulative error correction effect, we establish a digital simulation platform.