Robust Image Matching Methods Research Based On Optimized Conditional Sample Consensus

Image Matching refers to the process of identifying the point correspondences between two or multiple images of the same scene or project. This process is an important topic in the fields of image processing and computer vision, and it is also the basis of theories and applications in this field. To date, image matching has been widely used in many applications, including image registration, object recognition, object tracking, and camera calibration, and has played a decisive role in all these fields.Several methods have been proposed for image matching, and they can be devided into two categories: intensity-based and feature-based. These feature-based methods are more invariable to transformation, which can improve the stability and reliability of intensity-based methods. And they can also improve the efficiency of intensity-based methods. However, among data points detected by feature-based matching methods, there still exist abnormal data(also known as outliers), which are not consistent with the model. They have great interference on the robustness of model parameters. To refine the data and reduce the interference of outliers, we need to determine the constraint model of correspondences, and choose the appropriate robust method for model estimation. Then we can obtain the stable model parameters and improve the robustness of the matching results.This thesis applies fundamental matrix as a constraint model, which represents the epipolar geometry relationship of image pairs. Then it implements a Bay SAC(Bayes SAmpling Consensus) strategy, which pre-processes the data to obtain their prior inlier probabilities and choose the points with the highest ones for sampling. Based on the original Bay SAC algorithm, this paper contains the following works:I. It proposes two prior inlier probability determining methods: based on imaging geometry, and based on the statistical characteristic of parameters;II. It brodens the hypothesis testing method: introducing the method with likelihood estimation function;III. It simplifies the Bayes theorem for probabilities updating: considering point number consistent with models and likelihood estimation value, respectively.This thesis verifies the proposed methods by testing on several image pairs based on differrent platforms and stereo vision. The matching results indicate that no matter the stereo vision is known or not, the optimized Bay SAC algorithm can obtain more accurate inlier prior probabilities than original Bay SAC. The more outliers contain the pseudo-correspondences, the higher computational efficiency of our proposed algorithm gains compared with the original Bay SAC. Moreover, the optimized Bay SAC can overall achieve better matching accuracy than original Bay SAC.