Research On Parameter Calibration Method Of Monocular Camera For Ranging

As an important means of environment perception and distance perception of intelligent vehicles and unmanned vehicles,visual measurement is becoming a research hotspot in the field of computer vision.The ranging method based on monocular vision does not need image registration.Compared with the widely used binocular vision ranging,it has lower model complexity and application cost.The camera calibration method is an important factor affecting the preparation,real-time and accuracy of monocular vision ranging.However,the current camera calibration methods have the problems of complex calibration process and low integration with monocular vision ranging model,resulting in low ranging accuracy.In order to improve the accuracy of camera calibration and monocular vision ranging method,this thesis proposes a camera focal length calibration method based on monocular ranging model and a camera inclination calibration method based on artificial neural network and width measurement model.The complexity of camera calibration process in monocular vision oriented ranging model is effectively reduced and the ranging accuracy is improved.The specific work of this thesis is as follows:1.Firstly,based on the principle of linear imaging,the ranging model without camera inclination is extended to the ranging model with inclination.2.The camera focal length calibration method in the ranging model is improved.In this method,the focal length values including defocus and distortion at different imaging positions are regressed by polynomial to achieve the calibration of variable focal length.Compared with the traditional fixed focal length calibration,this method takes into account the influence of defocus on imaging,and there is no need to de distort and restore the pixel coordinates.3.This thesis makes an experimental study on the selection of data set capacity and training error threshold used to fit regression vector in the process of focal length calibration,and puts forward a scheme to control the test error by finding the minimum value of the ratio of test error to training error threshold.For the individual data whose training error is greater than the threshold,an error compensation algorithm for micro correction pixel coordinates is proposed.4.This thesis presents a camera inclination calibration method based on width measurement model and ANN model.Firstly,the method infers the width measurement model based on the ranging model to form the inclination constraint equation,and constructs a parallel ANN network to calibrate the camera inclination of two dimensions at the same time,which realizes the effect of calibrating the inclination of two cameras at the same time by a pair of feature points.The experimental results show that the ranging model with inclination in this thesis can measure the distance of ground objects in the field of view with high precision under any attitude of the camera.The focal length calibration and inclination calibration methods proposed in this thesis can effectively improve the ranging accuracy of monocular vision to more than 97%,and the average ranging accuracy can reach 99%.This shows that the ranging and calibration methods in this thesis have high research value for the application of intelligent vehicle positioning,obstacle avoidance,over-speed detection and so on.