Research On Feature Matching And Pose Estimation Algorithm Based On Monocular Vision

Visual positioning method,also known as visual registration method,is a branch of artificial intelligence.After the improvement of researchers from various countries,it has become a complete field.Visual positioning method has been widely used in robot path planning,augmented reality and virtual reality.In the case of unmanned driving,Baidu's unmanned vehicles put visual positioning technology and environment perception technologies combined to increase the ability of responding to complex environments(heavy traffic,poor road conditions,etc.).In short,visual positioning technology is slowly changing people's lives,but most visual positioning methods still have the problem of complex algorithms and high false-match rate.Therefore,in this article the matching algorithm and pose estimation algorithm are studied.In the VIM compiling environment that comes with Linux,the OpenCV vision library is used for implementation and verification.For the above problems,this paper designs the process of feature point matching and pose estimation based on monocular vision.The main research content of this paper is as follows:1.Analyze the three major coordinate systems and camera models.Explain the transformation between coordinate systems.Propose a correction model after distortion of the camera model.Apply image preprocessing to verify the advantages and disadvantages of graying.filtering,sharpening,and edge detection in preprocessing which are prerequisites for subsequent compilation programs.2.Analyze the image feature extraction and matching,and select the algorithm with high extraction quality in common feature points and feature corners respectively.The advantages of the algorithm are analyzed.Because of the large computation of the corner algorithm,the corner extraction algorithm is improved by combining the two algorithms to synthesize the fusion algorithm.Based on the Linux operating system,the program is compiled to determine the feasibility of the algorithm.3.Since the feature point registration algorithm still has limitations,this paper will add a random sampling agreement algorithm to filter the feature pairs processed by the fusion algorithm,reduce the false matching rate,and obtain an effective feature pair.4.In order to get good positioning result,this article,combined with the OpenCV open source visual image library and g2 o library under the Linux operating system,combines fusion algorithm with RANSAC,using the methods which named 2D-2D ?3D-2D and 3D-3D to accomplish the pose estimation.Experiments show that Surf-Harris fusion algorithm can effectively improve the efficiency of extracting feature pairs and effectively improve the accuracy of position information.