3D Reconstruction Based On Multiple View Geometry

The computer vision measurement system can be used to investigate the traffic accident later and reconstruct the scence. The 3D structure of the objects can be recovered from the correspondences of image pairs. However, most of computer vision systems can only get metric reconstruction, and the Euclidean reconstruction is necessary for our application. So staff gauge or calibration temple are placed on locale to achieve it in these systems which will affect the accuracy and bring inconvience. And if we use stereo system with constant focal, the Euclidean reconstruct can be obtained, while the measurement scope is limited. This debases the flexibility of the system, and the measurement accuracy and scope are all affected.In this paper, the 3D reconstruction from multiple views is researched to solve this problem. Based on the researching of 3D reconstruction and self-calibration, the technique which integrates the traditional calibration and self-calibration is applied to varying focal stereo rig.Firstly, the linear self-calibration using constant intrinsic parameter is researched. A novel stratified self-calibration method is presented in this paper to solve the problem that the image of absolute conic (IAC) or dual image of absolute conic (DIAC) must be positive definite for self-calibration techniques. Because the intrinsic can be estimated, so it is possible to avoid the positive definite problem with optimizing the intrinsic parameters directly. The genetic algorithm is used to get the intrinsic parameters which need not initial values. The performance of method is confirmed by a lot of synthetic and real images. The results show that the algorithm has good accuracy and is practicable.Common self-calibration use linear camera model, but the distortion from lens can not be ignored for high accuracy measurement. The grid template is used to calibrate the linear model generally. But this method is not usable if we use vary focal system. So a calibration method on line is presented in this paper. Hough transform is used to detect lines, and canny algorithm is applied to find edges and connect edges. Therefore, the difference between the edges and lines is the distortion which can help us rectify the distort image. Because of the need of system, the self-calibration on varying focal is researched. The Kruppa equation is deduced to calibrate the focal lengths from two views. The method is verified by theory and real test result. But the method is sensitivity to image noise, so the three or more images are advised.Based on the research of single camera system, the 3D reconstruction and self-calibration of stereo rig are also studied. A stereo rig system with varying focal cameras is presented which integrates traditional calibration and self-calibration. Principal points of the cameras and other parameters are calibrated by traditional method, and only focal length and distortion parameters are calibrated by self-calibration. And the transform matrix restriction is introduced to improve on images matching during this process. This method can depress the ratio of error match and get accurate transform matrix. This system combines the advantages of traditional calibration and self-calibration method and decrease the inconvenience for using template. Good accuracy and flexibleness can be acquired for traffic accident investigation.