基于卷积神经网络与扩展卡尔曼滤波的单目视觉惯性里程计  

Utilizing extended Kalman filter to improve convolutional neural networks based monocular visual-inertial odometry

作　　者：林立雄 郑佳春[1] 黄国辉 蔡国玮 Lin Lixiong;Zheng Jiachun;Huang Guohui;Cai Guowei(School of Information Engineering,Jimei University,Xiamen 361005,China;School of Mechanical Engineering and Automation,Fuzhou University,Fuzhou 350000;Fujian Chuanzheng Communications College,Fuzhou 350000,China)

机构地区：[1]集美大学海洋信息工程学院,厦门361005 [2]福州大学机械工程及自动化学院,福州350000 [3]福建船政交通职业学院,福州350000

出　　处：《仪器仪表学报》2021年第10期188-198,共11页Chinese Journal of Scientific Instrument

基　　金：福建省自然科学基金(2019J05024);国家自然科学基金(61803089)项目资助。

摘　　要：针对单目相机采集室外图像易受环境光照影响、尺度存在不确定性的缺点,以及利用神经网络进行位姿估计不准确的问题,提出一种基于卷积神经网络(CNN)与扩展卡尔曼滤波(EKF)的单目视觉惯性里程计。采用神经网络取代传统里程计中基于几何约束的视觉前端,将单目相机输出的估计值作为测量更新,并通过神经网络优化EKF的误差协方差。利用EKF融合CNN输出的单目相机位姿和惯性测量单元(IMU)数据,优化CNN的位姿估计,补偿相机尺度信息与IMU累计误差,实现无人系统运动位姿的更新和估计。相比于使用单目图像的深度学习算法Depth-VO-Feat,所提算法融合单目图像和IMU数据进行位姿估计,KITTI数据集中09序列的平动、转动误差分别减少45.4%、47.8%,10序列的平动、转动误差分别减少68.1%、43.4%。实验结果表明所提算法能进行更准确的位姿估计,验证了算法的准确性和可行性。The outdoor images collected by monocular camera are easily affected by the light intensity.The scale of images is ambiguous.In addition,the pose estimation of convolution neural networks(CNNs)is not accurate.To address these issues,a monocular vision-inertial odometry using CNN and the extended Kalman filter(EKF)is proposed.The CNN is used to replace the conventional odometry of the front-end vision based geometric constraints.The output of the monocular camera is used as the EKF measurement to correct the estimated pose of CNN.The error covariance of EKF is optimized by the CNN.The monocular camera pose data and the inertial measurement unit(IMU)data are fused in EKF to estimate the motion pose.The monocular scale informations and the cumulative errors of the IMU are compensated.Experimental results show that the proposed algorithm performs more precise pose estimation.The accuracy and feasibility of the algorithm are verified.Compared with the Depth-VO-Feat algorithm that relies on monocular images,the proposed algorithm combines monocular image and IMU data for pose estimation.The translation and rotation errors of the 09 sequence in KITTI dataset are reduced by 45.4%and 47.8%,respectively.The translation and rotation errors of 10 sequences are reduced by 68.1%and 43.4%,respectively.

关 键 词：视觉惯性里程计 神经网络 惯性测量单元 扩展卡尔曼滤波 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置] TH74[自动化与计算机技术—控制科学与工程]