Research On The Detection Algorithm For Catenary Geometry Based On Binocular Vision

With the rapid development of electrified railway, the detection technology of catenarygeometry has also changed. Because of the rapid development of computer technology,successfully visual detection technology is used to High-speed Rail inspection equipment. Butin the catenary geometry detection process, capture video data due to various factors, resultingin inaccurate detection results. Therefore, in the process of using catenary video detectionmethods, it is significant to research an image recognition method with robust and stronganti-interference ability. On the basis of analysis of catenary visual inspection system, thethesis makes indepth research on the key technology of camera calibration, the original imagepreprocessing and image feature extraction and etc in the system of measuring geometricparameters.First of all, according to the environment of parameter detection system, with the basicprinciples of camera calibration, the thesis uses coplanar method in camera calibration. In theprocess of particular calibration, under the consideration of the conditions of the camera lensdistortion, the thesis establishes nonlinear model and uses the least squares method for dataprocessing to solve the matrix parameters of the camera.Secondly, a new algorithm of image enhancement of contact based on Curvelet trans-form is proposed, for the problem of high contrast images collected by the camera. Accordingto the respective characteristics of high frequency subband and low frequency subband whichare obtained by curvelet transform, the low-frequency signal in the Curvelet domain isenhanced by the fractional differential. Meanwhile in the high-frequency field, on the edge ofthe high-frequency coefficients and noise are classified by the correlation of the Curveletcoefficient, an operator is proposed that the edge is enhanced and the noise is removed.Experimental results show that the proposed algorithm in this thesis has greatly improved interms of subjective and objective aspects.Finally, on the edge of the image features and feature point feature extraction algorithmare discussed. In the edge feature extraction, the stronger noise resistance edge detectionoperator that Smallest Univalue Segment Assimilating Nucleus (SUSAN) is used to detect thepantograph and contact wire edge. For this system, the algorithm is improved on the detectionaccuracy of the edge. In the point feature detection, Feature extraction method of catenarypantograph and the contact line of video image are studied in the complex background. ScaleInvarinat Feature Transform (SIFT) feature detection operator is selected as feature pointdetection. According to real-time deficiencies of the method, this thesis improves from twoaspects: one is to add the threshold T to the recent euclidean distance of feature points, thiscan avoid the calculation of high dimensional128poor square. Secondly, through simulation experiments founds that the number of feature points in the fourth group is almost no change,it can reduce the amount of computation by choosing the appropriate number of cut-off group.