Convenient Multi-Camera Self-Calibration Methods In Large Site

Along with the camera and computer cost reduction and the continuous development of related technologies, more and more multi-camera intelligent systems appear. As the foundation of these systems, camera calibration directly influences their performance. Compared with traditional calibration, multi-camera calibration algorithm in large site is more complicated. On one hand, this is because space position among different cameras is more complex, which means the complexity of field of view(FOV) of different cameras grows apparently. On the other hand, increasing number of cameras make calibration cumbersome. So it is hard to handle the calibration need under this site for most existing algorithms.This thesis takes a thorough research on the topic of multi-camera calibration in large site, especially for the site with only two views. A multicamera calibration algorithm based on essential matrix is proposed in this thesis. Besides, we have optimized the Self-calibration algorithm which based on the absolute conic, which have made the algorithm works with only two views while have kept the simplicity and accuracy. The main researches and contributions of this thesis include the following aspects:(1) Proposed and implemented an essential matrix. Based on multicamera self-calibration algorithm, the essential matrix is calculated through corresponding points, and then the camera's external matrix is restored by SVD. Then we calculated the scale factor by a 1D calibration object which consists of two infrared reflective balls connected by a stick with fixed length. Finally, we have taken the reciprocal of the effective corresponding pointnumber between the cameras as the weight, and used Dijkstra single-source shortest-path algorithm to select the optimal path to link the results.(2) Optimized and implemented a multi-camera self-calibration algorithm based on the absolute conic. The algorithm used SVD algorithm to decompose the scaled measure matrix, which is composed of the image coordinate and depth information of the corresponding point, into a projection reconstruction, and then calculated the homographic matrix, and finally restored the Euler reconstruction. But because of the mathematic model`s constraints, the current methods usually need more than three different views to complete the computation, this is hard to reach in large site with multi-camera. A new mathematic model is proposed which only needs two different views in this thesis based on the constant internal matrix. The model is strictly derived and proved.(3) Experiments have been designed to verify the correction and accuracy of these two algorithms. With results of the experiment, detail analysis has been given in this thesis, and it proved the accuracy, correctness and simplicity of the algorithms proposed in the thesis.