弱纹理环境下融合点线特征的双目视觉同步定位与建图  

Binocular vision SLAM with fused point and line features in weak texture environment

作　　者：龚坤 徐鑫 陈小庆 许悦雷 张兆祥 GONG Kun;XU Xin;CHEN Xiaoqing;XU Yuelei;ZHANG Zhaoxiang(Unmanned System Research Institute,Northwestern Polytechnical University,Xi′an 710072,China;National Innovation Institute of Defense Technology,Academy of Military Science,Beijing 100850,China)

机构地区：[1]西北工业大学无人系统技术研究院,陕西西安710072 [2]中国人民解放军军事科学院国防科技创新研究院,北京100850

出　　处：《光学精密工程》2024年第5期752-763,共12页Optics and Precision Engineering

基　　金：陕西省自然科学基金资助项目(No.D5110220135);中国高校产学研创新基金资助项目(No.2021ITA10022)。

摘　　要：针对室内弱纹理环境下基于点特征的视觉同步定位与建图(Simultaneous Localization and Mapping,SLAM)存在的轨迹漂移等问题,提出了一种融合点线特征的双目视觉SLAM系统,并对线特征的提取与匹配问题展开研究。为了提高线特征的质量,通过长度与梯度抑制、短线合并等方法,进一步改进LSD(Line Segment Detector)线特征提取方法。同时,通过将匹配问题转换为优化问题,并利用几何约束构建代价函数,提出了一种基于几何约束的快速线段三角化方法。实验结果表明,本文所提方法在多个数据集上的表现都优于基于描述子的传统方法,尤其在室内弱纹理场景下,其平均匹配精度达到91.67%,平均匹配时间仅需7.4 ms。基于此方法,双目视觉SLAM系统在弱纹理数据集上与已有算法ORBSLAM2,PL-SLAM的定位误差分别为1.24,7.49,3.67 m,定位精度优于现有算法。Addressing the challenge of trajectory drift in visual Simultaneous Localization and Mapping(SLAM)due to point features in texture-deficient indoor settings,this study introduces a binocular visual SLAM system that combines point and line features.It emphasizes the extraction and matching of line features within binocular visual SLAM.An enhanced line feature extraction technique,based on the Line Segment Detector(LSD)algorithm,is proposed.This includes improvements like length and gradient filtering,and the amalgamation of short lines.Additionally,the matching issue is redefined as an optimization challenge,creating a cost function based on geometric constraints.A novel,efficient line segment triangulation approach,leveraging the L1-norm sparse solution,is developed for effective line matching and trian‑gulation.Experimental evidence shows that our method surpasses traditional descriptor-based approaches across various datasets,especially in texture-sparse indoor areas,achieving a remarkable average matching accuracy of 91.67%and a swift average matching time of 7.4 ms.Employing this technique,our binocular visual SLAM system records positioning errors of 1.24,7.49,and 3.67 m on texture-sparse datasets,out-performing leading algorithms like ORBSLAM2 and PL-SLAM in positioning precision.

关 键 词：双目视觉 线特征提取 视觉同步定位与建图 特征匹配 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]