融合光度和深度的视觉里程计改进算法  

An improved algorithm of visual odometry combining intensity and depth

作　　者：黄宴委[1] 董文康 王俊[1] 陈少斌[1] HUANG Yanwei;DONG Wenkang;WANG Jun;CHEN Shaobin(College of Electrical Engineering and Automation,Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福州大学电气工程与自动化学院

出　　处：《福州大学学报（自然科学版）》2019年第6期746-752,共7页Journal of Fuzhou University(Natural Science Edition)

基　　金：福建省科技计划资助项目(2019H0007);现代精密测量与激光无损检测福建省高校重点实验室基金资助项目(2018XKA005)

摘　　要：针对视觉里程计中对纹理较少的环境难以获得高精度相机位姿的问题,提出一种融合光度和深度的视觉同步定位与绘图(SLAM)方法.首先构造基于光度和深度的相机位姿优化函数,通过t分布模型计算每个像素点的光度和深度误差权重,对每一帧图像引入帧权重系数λ来平衡光度和深度,并采用中值法求解帧权重λ,求得相机位姿.然后对10个国际标准数据集进行仿真实验,结果表明,对纹理丰富的环境,本方法能够保持DVO SLAM算法的建图精度;对纹理较少的环境,本方法的建图精度要高于DVO SLAM算法,绝对路径误差降低31.6%,相对位姿误差降低19.4%.Aiming at the problem that the high-precision camera pose is difficult to obtain in the environment with less texture in the visual odometry,this paper proposes a visual simultaneous location and mapping(SLAM)method that combines intensity and depth.Firstly,the camera pose optimization function based on intensity and depth is constructed.The intensity and depth error weights of each pixel are calculated by the t-distribution model,and the frame weight coefficientλis introduced for each frame image to balance the intensity and depth.The value method solves the frame weightλand obtains the camera pose.Simulation experiments on 10 international standard datasets show that the proposed method can maintain the accuracy of DVO SLAM algorithm for texture-rich environments.For environments with less texture,the accuracy of this method is higher.With the DVO SLAM algorithm,the absolute trajectory error(ATE)is reduced by 31.6%and the relative pose error(RPE)is reduced by 19.4%.

关 键 词：同步定位与绘图(SLAM) 视觉里程计 光度 深度 帧权重 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]