Research On Extraction And Matching Of Linear And Curvilinear Features In Vision Motion Analysis

Linear and curvilinear features are the key features for recognizing objects. And the two kinds of features are often used for intermediate medium in the visual motion analysis. So this paper concentrated on theories and methods how to extract the two kinds of features from the 2D image and how to match for the two kinds of features from image sequence.The methods based on edge detection and Hough Transform are the main ways for extracting linear and curvilinear features. This paper studied and discussed the advantages and disadvantages about the two kinds of methods. And then this paper analyzed basic theory of the Beamlet transform and its applications in linear and curvilinear feature extraction. Beamlet has its advantage in extracting linear and curvilinear features in image. But the traditional Beamlet Transform has some problems: Firstly, it is defined on continuous function domain. When this transform is applied to image processing, they have to use interpolation on the discrete image to get a continuous function, and then take the transform on the continuous function. That process is complicated and cost much time. Secondly, the algorithm based on the BD-RDP that uses the complexity-penalized energy function in extracting features often lost small lines, especially in extracting complicated real image. Therefore, a modified algorithm was presented which is faster and can extract more linear and curvilinear features. It classifies Beamlets by the beamlets'geometric properties, defines a new quantity and establishes a new energy function. Experimental results demonstrated the reduction of the computation and the effectiveness in the complicated real images.The method of feature-based depended on the correspondence problem between features very much in determining 3D structure and motion. Correspondence problem is also called matching problem. The matching problem is closely related to the representation. This paper researched many methods about line feature and curve feature. A line feature matching technique based on eigenvector approach and a curve representation method based on included angle chain were introduced in detail in the paper. The number of candidate models, which derive from Line feature Matching Approach based Eigenvector, is still large. Aiming at the defect, an improved approach was presented. It reduces the number of candidate models and calculation of matching, and the speed of matching is also improved. Experimental results verified the performance of the proposed approach.Finally, a new method for curve representation and matching was presented, which was named"Beamlet Included Angle Chain—BIAC". It utilizes a multiscale structure—beamlets—that is designed primarily for linear and curvilinear features. It also utilizes method of the Included Angle Chain. The new representation is invariant to rotation, scaling and translation. And the method is insensitive to disturbances. Experimental results and algorithm analysis demonstrated the reduction of the calculation and the effectiveness and rationality for curve matching.