Research On Camera System Calibration Method For Pose Estimation

Recently the application of robot on the industrial production line has been paid close attention. The robot system put forward the challenge that the flexible production line need to adapt to the production of different industrial products according to the demand. The combination of computer vision and robot technology improves the adaptability of industrial production line greatly. It is an effective way to improve the ability of the robot by using the binocular vision system. In this system, the researches on high precision camera calibration and pose estimation methods are two key points. The main innovations and research results of this thesis are shown as follows:One of the main factors that influence the imaging quality is the lens distortion. In order to improve the accuracy of distortion correction, a vanishing point reprojection model is proposed. This model improves the accuracy of the line equation estimation on the distorted image effectively. Due to the parameters coupling, the nonlinear optimization results would fall into the local optimal. Therefore, recursive individual optimization model is proposed to decouple the parameters. Based on these two models this thesis proposes a vanishing point reprojection distortion correction method. The simulation experiments indicate that the correction under the noise circumstance has high accuracy. The results show that the correction accuracy is 1.5 pixels, and it degrades smoothly as the noise increases. It indicates that this algorithm has anti-noise ability. Then, real data experiments verify that this method can correct distortion effectively.Camera calibration is an important step for getting object space coordinates in binocular stereo vision system. In order to improve the accuracy of camera calibration, a method based on varnishing point consistence constraint model is proposed. This model is composed of the distortion correction method and the homography solving method, which improves the accuracy of varnishing point location. Thus, the precision of camera calibration is improved by this model. Simulated experiments show that the method works well under the noise circumstance, the average reprojection error is 0.04 pixels. It indicates that the method is able to achieve high reprojection acurracy on highly distorted image. Then, in real data experiments, the maximum reprojection errors are 0.60 pixels and 0.50 pixels respectively.The object pose estimation is a fundamental step for robot fetching. In order to improve the applicability of the pose estimation algorithm, a method based on a remote point pair elimination strategy is proposed. Then, in order to reduce the influence of the initial pose on the result of the traditional iterative closest point method, a hybrid of simplex method and simulated annealing is introduced which searches the optimal solution in the paramter space. The remote point pair elimination strategy is used to improve the accuracy of the pose estimation, which is realized by discarding the mismatching point pairs with large distance between them. The experiment results show that the proposed method is better than the tranditional ICP method.This thesis designed a binocular stereo vision system for fetching object. A multi-task software architecture based on DSP/BIOS is proposed, and the pose estimation algorithm is operated in this system successfully. In experiment system, the KUKA robot is used for providing accurate object pose data. Then, the precision of the pose estimation algorithm in this thesis is evaluated by comparing the results with the true object pose data. The experiment results show that this method is effective for fetching object, and the total location error is 2.7mm.