Research On Monocular Depth Estimation Method Based On Multi Focus Disparity

Monocular depth perception is a hot research direction in machine vision.This thesis proposes a monocular depth estimation method based on multi focus parallax,and studies the performance of camera calibration methods,multi focus image feature extraction and matching methods,and multi focus parallax depth estimation.The specific work content and research conclusions are as follows:(1)Studied the relevant principles of multi focus disparity monocular depth estimation methods.Firstly,the principle of multifocal image imaging is studied,and the sources of multifocal parallax are analyzed;Secondly,obtain the geometric relationship between the depth of the object point and the corresponding image pixels,as well as the parameters such as image distance and principal point coordinates,and establish a multi focus disparity monocular depth estimation model;Then,based on the multi focus disparity monocular depth estimation formula and the requirements for image feature coordinates and camera parameters,classic feature extraction and matching methods and camera calibration methods were analyzed,laying a theoretical foundation for the study of multi focus image feature extraction and matching and camera calibration methods based on single images in this thesis.(2)A camera calibration method based on a single image was proposed.Firstly,based on the linear features of the calibration board,an objective optimization function was constructed to solve the radial distortion coefficient and principal point coordinates of the lens,transforming the nonlinear calibration problem into a linear problem;Then,based on the camera model,the focal length and external parameter matrix are linearly solved.The experimental results show that the reprojection error of the proposed method is 0.21～0.24 pixels,which has higher accuracy and robustness compared to the existing single image calibration methods of2.68～2.85 pixels,laying the foundation for the calculation of multi focus disparity.(3)A multi focus image feature extraction and matching method based on optical flow tracking is proposed.Firstly,conventional image feature extraction methods are used to extract the features of the focused and clear areas in multi focus images;By utilizing the optical flow tracking algorithm,the features extracted by conventional methods are mapped to the defocused blurry areas of other multifocal images,avoiding the uncertainty of existing methods directly extracting features from the defocused blurry areas.Compared with existing methods,the proposed method can effectively extract features from defocused blurry areas without losing the number of features in the focused clear area.The number of features in blurry areas is more than 2.25 times that of the comparison method,and the total number of features is significantly increased,with a small error of only 0.11～0.38 pixels,which has higher accuracy compared to the comparison method.In addition,the proposed method effectively completes image feature matching while extracting features,obtaining multi focus disparity,laying the foundation for subsequent multi focus disparity monocular depth estimation.(4)Analyzed the error of multi focus disparity monocular depth estimation method.Firstly,theoretical error analysis of the depth estimation model is conducted using simulation experiments,including estimating the source and magnitude of errors for each parameter in the depth estimation formula,as well as analyzing the depth estimation errors of each parameter at different values;Then,actual depth estimation experiments were conducted based on simulation results,verifying the feasibility of the proposed depth estimation method and the accuracy of theoretical error analysis.