Author(s): |
Shulin Sui, Institute of Autonomous Navigation and Intelligent Control, Qingdao University of Science &Technology Qingdao, china Wei Zhao, Institute of Autonomous Navigation and Intelligent Control, Qingdao University of Science &Technology Qingdao, china Wei Shao, Institute of Autonomous Navigation and Intelligent Control, Qingdao University of Science &Technology Qingdao, china Chaoyang Wang, College of Information Science and Engineering, Shandong University of Science and technology Qingdao, china |
Abstract: |
Camera calibration is a primary and crucial task in the field of computer vision. The previous methods, which are calibrated by neural network, are commonly calibrated implicitly using neural net- work by means of its ability to fit the complicated nonlinear mapping relation. In this paper using BP neural network, the projection matrix between the world 3D world points and the corresponding 2D im- age pixel points can be obtained, Starting from random initial weights, the net can get camera extrinsic and intrinsic parameters and make the rotation matrix satisfy the orthogonality constraints by the train of variable learning rate and momentum vector algorithm. The advantage of this method is verified with the simulation of both synthetic data under different noise conditions and real images. Compared with other traditional calibration methods, experimental results show the calibration approach based on BP neural network is very efficient and robust, and the accuracy is good enough.
|