Research On Image Matching And Tracking

Image tracking has wide application in military target tracking, human-machine interface, driver assistant system and so on. The existing image tracking algorithms can be categorized into four classes: model matching based method, motion information based method, filtering based method and classification based method. Among which, the model matching based mehod is simplest, which is easy to manipulate and has widest application. There are mainly three key elements in model matching based image tracking method, which are representation of image features, similarity measure and target search method respectively. This dissertation made a thorough research on the second element and some useful exploration in the first and third elements. The contributions of the dissertation mainly include the following aspects:Firstly, a mathematical model of image matching is developed for matching performance evaluation of similarity measures, based on which, the matching performance of several often used similarity measures has been analyzed and compared. There are many similarity measures in image matching field, the matching performance of which can only be evaluated by experiments so far. In order to evaluate the performance in theory, through the analysis of the essential reason of matching failures, an image matching model with position variable is constructed on basis of image feature classification. The model can be employed to evaluate the matching performance of different similarity measures, direct us to select or improve the existing similarity measures and develop new similarity measure. Based on the proposed matching model, several often used similarity measures have been analyzed and compared in theory. The critical conditions of the presence of matching bias and mistake are given. Actual experiments tested the correctness of the analysis results and validated the effectiveness of the matching model.Secondly, with the direction of the matching model, a modified Bhattacharyya coefficient is develped, which greatly improves the matching performance compared to the original Bhattacharyya coefficient. What is more, a new similarity measure called Background Suppressed similarity measure is also developed. It has been demonstrated by the matching model that the Background Suppressed similarity measure is linear with respect to the matching position. It can effectively reduce the influence from the background feature included in the target candidate. Experimental results showed that the Background Suppressed similarity measure can greatly improve the target tracking accuracy, which validated the matching model from another point of view.Thirdly, a Posterior Probability measure is proposed, based on which two fast image tracking algorithms are developed. It is proven that with specific parameter, Background Suppressed similarity measure is equivalent to the posterior probability of the matched feature. Based on which, Posterior Probability measure is developed. The Posterior Probability measure can effectively suppress the influence from the background by introducing the statistical feature of the search region, and meanwhile up-weight the real target feature. Comparing with Bhattacharyya coefficient, the peak modality of the Posterior Probability measure is improved greatly. Another benefit of Posterior Probability measure is the low computational complexity and obvious global maximum. Through analysis, it can be seen that the Posterior Probability measure can be obtained as the summation of pixel similarities. Based on which, two fast tracking methods have been developed, namely Fast Translation Algorithm and Centroid Iteration Algorithm. In addition, a multi-resolutional and scale adaptation scheme is also developed. Quantities of comparison experiments showed the effectiveness of the above tracking algorithms.Lastly, based on the research of the important image feature, i.e. corner, a new Accumulative Intersection Space based corner detection method is developed. Corner does not have explicit formulation with parameters, so the troditional Hough transform can not be employed to transform the corner detection into maximum search in parameter space. A new random corner detection method under Monte Carlo scheme is developed, which transforms corner detection into local maximum search in the accumulative intersection space, not the traditional parameter space. Accumulative intersection space is a new concept which is proposed according to the reality that corner is actually the intersection of lines. The proposed algorithm is demonstrated in the dissertation. The steps of the algorithm are also given. The new algorithm is isotropic, robuts to rotation and insensitive to noise. It can also avoid false corners on the diagonal edge effectively. Plenty of experiments showed that, the new algorithm is superior to Harris algorithm, Shen & Wang algorithm, Wang & Brady algorithm and SIFT feature in some aspects, which can also be employed in image tracking successfully.The dissertation has made some thorough research on several key problems in image tracking field. Especially, a theoretic evaluation system of similarity measure has been developed, which supplies theoretic foundation for this research field.