融合自监督单目图像深度估计的视觉里程计  

Visual Odometry with Self-Supervised Monocular Depth Estimation

作　　者：成立明 陈建新[1] 陈瑞[1] 陈志敏 CHENG Liming;CHEN Jianxin;CHEN Rui;CHEN Zhimin(College of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,南京210003

出　　处：《计算机工程与应用》2021年第15期230-236,共7页Computer Engineering and Applications

基　　金：国家自然科学基金(61901227);南京邮电大学项目(NY217021,NY218014)。

摘　　要：针对直接法DSO(Direct Sparse Odometry)存在的明显的尺度不确定性问题,对尺度不确定性给系统定位精度带来的影响进行分析,提出将对单幅图像进行深度估计的深度学习网络和DSO相结合的融合算法;针对DSO后端耗时问题,提出运用预处理共轭梯度(Preconditioned Conjugate gradient,PCG)算法优化后端求解部分。在KITTI公开数据集上与ORB-SLAM2、DSO、LDSO进行对比测试,系统的定位精度得到显著提高。Aiming at the obvious scale uncertainty problem of the direct method DSO(Direct Sparse Odometry),the impact of scale uncertainty on the system positioning accuracy is analyzed,and using a deep learning network that performs depth prediction on a single image is proposed.For the time-consuming problem of the back-end,a Preconditioned Conjugate Gradient(PCG)algorithm is utilized to accelerate the back-end solution.Compared with ORB-SLAM2,DSO,LDSO on the KITTI public dataset,the accuracy of localization of the system is significantly improved.

关 键 词：深度估计 视觉里程计 尺度 预处理共轭梯度 定位 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]