点线特征自适应融合室内SLAM算法  

Point-line Feature Adaptive Fusion Indoor SLAM Algorithm

作　　者：刘少哲 刘作军 胡超芳 陈海永 LIU Shao-zhe;LIU Zuo-jun;HU Chao-fang;CHEN Hai-yong(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300130,China;School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300130 [2]天津大学电气自动化与信息工程学院,天津300072

出　　处：《小型微型计算机系统》2023年第5期1015-1022,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61941303)资助;天津市科技计划项目(19YFHBQY00040)资助。

摘　　要：传统的视觉同时定位与地图创建(SLAM)依赖于点特征来估计相机位姿.然而在室内环境中存在大量低纹理区域,使得提取足够多的点特征变得困难.此外,当相机抖动剧烈或转向过快时,基于点特征的SLAM系统也并不鲁棒.针对上述问题,本文提出了一种基于RGB-D的点线特征融合SLAM算法,利用点特征和线特征的优点,在困难环境下获得了鲁棒的结果.首先,提出了一种基于特征丰富度的特征提取策略.解决在模糊和低纹理区域内提取特征困难的问题.其次,设计了一种点线特征关联图,优化线特征匹配效果.该方法不仅参考了线特征之间的相似关系,还考虑了点线特征之间的几何关系.最后,在构建光束法平差的成本函数时建立自适应模型,实现点线双模态特征的"无缝融合".本文分别在两个公开数据集和室内真实场景中进行了算法评估,并与其他先进算法对比.结果表明本文提出的算法具有更好的整体性能.Traditional visual simultaneous localization and mapping(SLAM)depends on point features to estimate camera pose.However,a large number of low texture areas exist in the indoor environment,which makes it difficult to extract enough point features.In addition,the SLAM system based on point feature is also not robust when the camera jitters sharply or turns too fast.To solve the above problems,this paper proposes a point-line feature fusion SLAM algorithm based on RGB-D.The advantages of point feature and line feature are utilized,and robust results are obtained in difficult environments.Firstly,a feature extraction strategy based on feature richness is proposed.It solves the problem of less extracting features in blurred and low texture regions.Secondly,the point-line feature association graph is designed to optimize the matching effect of line features.This method not only considers the matching relationship between line features,but also refers to the geometric relationship between point-line features.Finally,when constructing the cost function of Bundle Adjustment(BA)optimization method,an adaptive model is established to realize the“seamless fusion”of point-line dual-modal feature.In the experiment,this SLAM algorithm is evaluated on two public datasets and indoor real scenes.Compared with other advanced algorithms,the results show that this SLAM algorithm has better performance.

关 键 词：机器视觉 同时定位与地图创建 点线特征 自适应模型 低纹理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]