Study On Camera Calibration Methods

As the modern computer technology growing fastly, an emerging technology, Computer Vision, has gradually formed. It mainly uses computer to extract informations from videos or photos for simulating people’s visual capabilities, then processes and understanding them, at last, these informations can be used for real-time monitoring and controlling. Intelligent visual surveillance is one of the most important applications in Computer Vision, it uses cameras controlled by computer to replace people, undertaking the monitoring tasks. It analysises and processes a sequence of video frames to implement monitoring. However, camera calibration is a primary and vital topic in intelligent visual surveillance.This paper introduces the background and basic theory of camera calibration, then discusses three methods of camera calibration and makes related experiments.The main job of this paper is as follows: firstly, we introduce the exsisting calibration approach based on2D calibration pattern, it gets the camera intrinsic and extrinsic parameters by the property of vanishing point, then optimizes and refines the intial values by Guass-Newton. After that, for the problem of the above method can not calibrating some special kind of pictures, we propose a new calibration algorithm to improve the above-mentioned approach, according to the results of the experiments, the new method can overcome the problem which the old one has. Secondly, because the classic non-linear camera calibration method has computational complexity and poor convergence and is Easy to fall into local optimum and can not guarantee the rotation matrix to be the orthogonal matrix, we propose a two-stage calibration method based on Genetic algorithm, and design the algorithm flow, the experiments verify that this method can solve the problems which mentioned above, and is more accurate than the method in chapter3. Thirdly, the requirement of the camera calibration in the intelligent visual surveillance is that the calibration object is as simple and common as possible and the calibration method is as fast as possible, but the calibration approach based on2D calibration pattern can not meet the above demands, so for this problem, we present a new calibration method by using vertical ground objects (human ect.). This approach gets vertical vanishing points and the plane vanishing line by top-points and bottom-points of vertical objects. Then we get the camera intrinsic and extrinsic parameters using the obove data. As the accuracy of vertical vanishing points and the plane vanishing line is of important, we use singular value decomposition (SVD) and fitting refinement method to guarantee its accuracy, which is also robust.In the last, we test the application of the calibration method by using vertical ground objects by experiments, which is used to measure the real3D distance by videos or photos. From the experiments, we can testify the robustness and precision of proposed calibration method.