| Machine vision is a quick-developing domain. The research focus was changing from test in laboratory to application in practice in 1980s. One of its important purpose is to recover 3D coordinates from 2D images, then to reconstruct 3D scene. This is applied widely, such as robot navigation, reverse engineering, object recognition, virtual reality and antique reparation, etc. Generally speaking, the reconstruction task of scene can be finished through two different ways: passive one and active one. The stereo rig is a typical passive way. Its basic principle is parallex. It is usually consisted of two cameras. The intrinsic and extrinsic parameters of the cameras must be pre-calibrated accurately before work. All the parameters must be kept constant during application. However, the object structure or scene surface may be variable on many practical cases, which requires to modify the parameters of vision system, for example, focus length, aperture and obliquity. The system will can not work exactly if the modification is not done. On the other hand, the modification change the original parameters and it is difficult to re-calibrate the system while it is working. In order to resolve ths problem, researchers want to avoid pre-calibration with precise target. They hope to recover the parametes of camera by virtue of geometry information of scene. Thus, the self-calibration of vision system is coming to be a focus in machine vision.Maybank and Faugeras firstly proposed self-calibration theory in 1992. Then, manyresearchers presented successively some effective self-calibration/uncalibrationmethods to reconstruct 3D scene. Whereas, most of their research obey thefollowing assumption: A series of images of the same scene captured from differentviewpoints are needed to reconstruct its 3D geometrical structure and the intrinsicparameters must be constant. This method is another typical passive one. It seems tobe easy for application, but in practice it contains the following problems: (1) Atleast 2 images are required, which makes it unfitted for variable or dynamic scene;(2) It suffers from the ambiguity of correspondence between the camera images, especiallythe images of free-form surface or natural scene, which is difficult problem to solve till now. (3)The reconstruction is equal up to a transformation matrix or a scale factor, it is not a realEuclidean result but a relative result or similar reconstruction. It can not obtain the absolutedistance between points.To avoid the problems mentioned above and to achieve real Euclidean reconstruction, active vision can be adopted. Structured light or pattern projection system can be used for this purpose, such as scanner with dot or line laser, encodedstructured light, etc. Although the scanning method can solve the correspondence problem, it must be pre-calibrated accurately using precise target. Many research work and useful methods have been done on this method, but they can only be applied on some special cases. It is not easy to implement on-line and real-time calibration for the vision system. It is almost impossible to reconstruct 3D scene from single image. As a result, the active vision system with encoded structured light illumination is most effective way to achieve real Euclidean reconstruction of 3D scene, especially for varible and dynamic scene, based on single image. Therefore, it is very significant to study what kind of projection pattern can be adopted and how to characterize the scene surface so that 3D Euclidean reconstruction can be acquired based on single image by uncalibrated camera. What's more, the parameters of camera can be modified following the variance of scene.This paper mainly focuses on the production of encoded structured light and flexible uncalibrated 3D Euclidean reconstruction technology. An self-adaptive Euclidean reconstruction method is presented based on single encoded image. It can recover the 3D shape of scene by a uncalibrated camera whose parameters are allowed to change...
|
PDF Full Text Request