可变精度的神经网络摄像机标定法

Variable precision camera calibration using neural network

提出了一种提高摄像机标定精度的方法。通过摄像机径向畸变模型,建立根据畸变严重程度自动改变区域划分数目的方法,对远离图像中心畸变程度严重的区域,划分细密;而靠近图像中心畸变轻微的区域,划分粗疏。通过对摄像机径向畸变区域进行划分,并且对每个畸变区域的像素进行单独的处理,构造相应的神经网络,得到整个畸变区域的处理结果,并对于不同的划分结果进行比较分析。分析比较得出:采用可变精度的神经网络摄像机标定法,可以大幅度提高标定的精度,划分数目越多,标定的精度越高,实验中识别率最高可达到99.45%.

A method is proposed to improve the precision of camera calibration. Based on the radial distortion model of a camera, the approach is presented, in which the partition number of distortion region can be adjusted automatically. In a region far from the image center, where distortion is high, the number of partition is big. While in the region near the image center, where the distortion is low, the number is small. Through the partition of the camera distortion regions and processing pixels in the corresponding regions, the neural network can be built. Then, the calibration result can be obtained. The processed results of the novel method were compared with different partitions. The conclusion is that the number of partition is bigger and the calibration precision is higher. The precision can reach 99.45% at the maximal partitions.