Repeatability Evaluation And Precision Location Method Of Line Feature For Image Matching

Feature extraction has great significance in the field of imagematching、object description and recognition、motion estimation、 targettracking and so on. The line segment is the "edge" and the "line" of image."Edge" can be defined as the dividing line between the local areas with different characteristics, and the "line" is a line formed with a small edge. For the line features in digital image, it is important to evaluate and locate them.Recently, there were many papers about image matching based on the line segment.But, most of them focus on the matching algorithm, and doesn’t concern the line segment detector. How to take a good line detect algorithm for image matching, and how to improve the precison of line segment is not mentioned. Based on this question, this thesis aims at the evaluation and precison location method of line segment detector. And several meaningful conclusions and algorithms are deduced through analysis and investigation. The main work of this thesis can be summarized as follows:1、Introduce four kinds of line segment detector algorithm and do a comprehensive study on these detectors. Propose a criterion for evaluating line segment detector:repeatability. And the criterion directly measures the quality of the feature for tasks like image matching, object recognition and3D reconstruction.2、Employ homography to describe the map relationship between the segments detected in stereo pair images. Propose a criterion for judging the corresponding line segment based on the geometric relationship. Based on the proposed criterion, repeatability of several typical line detection methods are evaluated by using benchermark stereo pair images.3、In recently many line segment detectors are proposed, but the line segment also has an integer pixel precision. So how to extract lines with sub-pixel precison is important for high-precison3D reconstruction.In this paper, a line segment detector with sub-pixel precision is proposed based on the strength of Harris. Firstly use the embedded confidence edge detection algorithm (referred to as the EDISON algorithm) to extract the edge, and refine the edges to single pixel. Secondly, fit each pixel in its normal direction using a parabola with the strength of Harris to abtain edge with sub-pixel precison. Thirdly, split the edge with minimue error and fit the edge with the least square to obtain a sub-pixel line segment.The theory and algorithm of the presented methods are implemented using Visual C++6.0and MATLAB7.8, and benchermark stereo pair images as well as real close-range images and aerial images are employed for experiment analysis. The experiment results show that the line segment extracted is well for image matching、3D reconstruction and camera calibration.