The Research On Camera Calibration Technology Based On Multi-view Images

With the advances of humans'research on 3D image technology and its application, multi-view image technology acquires great development. Multi-view image technology enabled computers the capability of visual cognition as human beings, which makes it can cognitive environment information through two dimension image. As one of the important technologies of multi-view images, camera calibration has also been developed rapidly. As camera calibration is the basis of some technologies, such as three-dimensional image reconstruction and camera tracking, the research on camera calibration is significant.There are three kinds of camera calibration method, calibration object-based method, active vision-based method and the camera self-calibration method. Calibration object- based method can get higher accuracy, but the calibration process is time-consuming. Therefore, calibration object- based method can not be used for online calibration and the situation without calibration block. In this kind of method, Zhang Zhengyou's method is considered as the classic camera calibration method, which is often used as the standard of camera calibration. The advantage of active vision-based method is simple and it can often get linear solution. However, when the camera motion pattern is unknown, this method is often useless. In view of the above consideration, this paper put emphasis on the camera self-calibration method.First, this paper studies stereo matching technology, which is one of the important part of camera calibration. This technology is divided into two categories, namely block-based stereo matching technology and feature-based stereo matching technology, in which block-based stereo matching method is simple, fast and feature-based stereo matching technology is relatively complex, time-consuming, but the matching accuracy is higher than that of block-based stereo matching. By comparison, this paper chooses the feature-based stereo matching technique as the stereo matching part of camera calibration.Nonlinear fitting technique is another important component of camera calibration. In this paper, two nonlinear fitting technologies, genetic algorithms and LM algorithm, are studied. Among them, the genetic algorithm has strong global search ability and does not need a given initial value, but it is weak in local search and its constraint functions must converge to the minimum error; LM algorithm has rapid convergence properties of Gauss-Newton method, but it needs the initial value close to the optimal solution. This paper proposed a method to estimate the camera intrinsic parameters, which combined genetic algorithm and LM algorithm. The method combines the global search properties of genetic algorithm and loc al convergence properties of LM algorithm.Based on the above research and analysis, this paper proposed a camera self-calibration method, which is based on multi-view images of single scene. In this method, multi-view images are processed through the following steps in order, that is feature matching, establishing the matching feature points set, select the three frames which are most suitable for calibration. Finally, the genetic algorithm and LM algorithm are combined to estimate the camera intrinsic parameters. Experimental results show that the accuracy of the result get from this method is higher. When compared with the method of genetic algorithm, the results are closer to the results of Zhang Zhengyou. Furthermore, the method proposed in this paper has the advantages of common camera self-calibration method, which do not need any calibration object and special moments of camera. Therefore, when compared to Zhang ZhengYou's method and some other traditional calibration methods, the proposed method in this paper has higher theory value and practical value.