Research On Feature Matching And Target Tracking Algorithms

Computer vision is a very important field in computer science, because it consititutes an essentail part of human-sensory system and has tramandous applications. Artificial intelligence is one of such applications, in which computer vision has been utilized as pattern recognition, 3D reconstruction and object tracking, etc. Therefore, computer vision has been regarded as high-demanding as genetic engineering and is becoming one of two important science technologies on the development of human society.Given a pair of images, how to detect and match features is one of the key technologies in the field of computer vision. For instance, in an intelligent decision systems for a dynamic environment, in order to analyze image sequences from the visual system hence instructions can be given to achieve the task of object tracking, images are required to be detected and matched, which is the main subject of this work. In particular, this paper studies image feature-matching, and its application on object-tracking system, mainly including two parts:Firstly, according to lieterature works on multi-constraints feature matching, by programming, we have realized the feature matching, which has feature detection, description, matching and matches filtering. However, we found that there are some limitations in this mechod, such as difficulty in threshold selectrion, small quantity in matches, etc.In order to overcome the above disadvantages, we propsoed a matching algorithm based on epipolar insert image features. In such, we have studied the epipolar geometry and introduced the epipolar for images, significant improvement was achieved.To futher improved the performance, a new method based on SURF feature-matching via epipolar constraint was proposed, in which RANSAC can be obtained by tightening parameters fundamental matrix with high quality. The process and implementation of this method is simple and accurate, which is able to enhance the number of correct matches and handle different type of images.Secondly, to utilize our result to solve problem in computer vision, we have built a system of object tracking which is consensus-based matching and tracking. The process is used to achieve the desired effect of object tracking, including feature matching, feature tracking, modeling, learning and updating. The experimental results showed that the method had advantages of tracking accuracy, timeliness, regain lost and scale change.We also have studied a comparative method of high-speed tracking with kernelized correlation filters. The method can track fast and accurately, but it has no features of scale change and regain lost.