宽视角相机相对姿态测量方法研究  

Research on relative attitude measurement method of wide-FOV camera

作　　者：李咸静 王鉴[1,2] 郭锦铭 张璇[1,2] 韩焱 Li Xianjing;Wang Jian;Guo Jinming;Zhang Xuan;Han Yan(Shanxi Key Laboratory of Signal Capturing and Processing,North University of China,Taiyuan 030051,China;School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学山西省信息探测与处理重点实验室,太原030051 [2]中北大学信息与通信工程学院,太原030051

出　　处：《电子测量技术》2022年第4期107-113,共7页Electronic Measurement Technology

基　　金：国家自然科学基金(61801437,61871351,61971381);山西省研究生创新项目(2021Y609)资助。

摘　　要：针对目前相机姿态估计方法都存在视觉局限性的问题,使用鱼眼镜头作为视觉传感器进行姿态估计,进而实现宽视角相机相对姿态估计。但鱼眼成像在具有宽视场优点的同时,伴随着严重的非线性畸变导致其在不同的方位和距离下具有不同畸变扩散的问题,为此提出了一种直接利用鱼眼图像的非线性进行相机相对姿态测量的方法。首先,构建鱼眼数据集kitti_FE;其次,使用卷积神经网络进行特征提取后结合长短时记忆网络进行双向循环训练,实现相机相对姿态的端对端输出;最后利用迁移学习的方法对实际场景进行相机姿态估计。为了验证所提方法的鲁棒性和精确度,在相同实际场景下,利用所提方法分别与现有框架CNN、DeepVO和CNN-LSTM-VO-cons进行对比。实验表明,该方法分别比现有框架的相机姿态估计精度提高了32%、29%和25%,而且在高速运动下该方法更具有稳定性。In view of the visual limitations of the current camera attitude estimation methods.In order to realize the relative attitude estimation of wide-FOV(fields of view camera),this paper used fisheye lens as visual sensor for attitude estimation.While fisheye imaging has the advantages of a wide-FOV,it is accompanied by serious nonlinear distortions,which leads to the problem of different distortion diffusion at different azimuths and distances.Therefore,this paper proposed a method to directly use the non-linear characteristics of the fisheye image to measure the relative pose of the camera.First,established the fisheye dataset kitti_FE.Secondly,used convolutional neural network for feature extraction and then combined with long short-term memory network for bidirectional loop training to achieve the end-to-end output of the relative posture of the camera.Finally,the method of transfer learning was used to estimate the pose of the fisheye camera in the actual scene.Experiments show that the proposed method is 32%、29%and 25%higher than the camera pose estimation accuracy under the existing frameworks of CNN,DeepVO and CNN-LSTM-VO-cons,respectively,and the proposed method is more stable under high-speed motion.

关 键 词：宽视角运动姿态估计 非线性光学 鱼眼成像模型 深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP277[自动化与计算机技术—控制科学与工程]