School of Mechanical and Electrical Engineering, Chengdu University of Technology, Chengdu 610059, China
Show less
文章历史+
Received
Published
Revised
2025-01-16
2028-07-28
2025-10-21
Online Date
2026-09-20
PDF (17033K)
摘要
目前多数视觉即时定位与地图构建(simultaneous localization and mapping, SLAM)方案都是通过提取环境中的特征点来估计位姿,在纹理较少的弱纹理环境中仍存在较大的局限性。为此,在SLAM系统中引入线特征以保证系统能在弱纹理场景中稳定运行。但目前融合点线的视觉SLAM方案存在实时性和精度不足的问题,因此提出基于改进点线特征融合的的视觉惯性SLAM算法。算法前端中,采用FAST(features from accelerated segment test)角点作为特征点提取算法,对ELSED(enhanced line segment detection)算法进行增加短线合并、梯度阈值参数调整,并将四叉树均匀化分布特征点扩展到点线特征,提出改进的点线特征提取算法,减少高纹理区域和特征分布不均的情况对系统精度的影响。对点线特征的跟踪,均采用改进型光流法追踪,将惯性测量单元(inertial measured unit, IMU)得到的位姿信息和已知的特征点深度计算光流法的初值,代替原本的图像金字塔迭代过程,从而节省计算资源,满足系统的实时性。最后,在实际场景中将该系统与优秀的开源方案进行实验对比,验证了所提算法的实时性和精确性。实验表明,本算法可为工业巡检、仓储物流等场景下的机器人提供高鲁棒性定位解决方案,具有显著的产业应用前景。
Abstract
Most current visual simultaneous localization and mapping (SLAM) solutions are based on the extraction of feature points for pose estimation, which still presents significant limitations in weakly textured environments. To address this, line features were introduced into the SLAM system to ensure stable operation in such low-texture scenes. However, current point-line fusion based visual SLAM schemes were often challenged by insufficient real-time performance and accuracy. Therefore, an improved visual-inertial SLAM algorithm based on point-line feature fusion was proposed.In the front-end of the algorithm, FAST (features from accelerated segment test) corners were employed as the feature point extraction method. The ELSED (enhanced line segment detection) algorithm was enhanced through short-segment merging and gradient threshold parameter adjustment. Furthermore, the quadtree-based method for uniform feature point distribution was extended to both points and lines, resulting in an improved point-line feature extraction algorithm. This improvement was designed to reduce the impact of high-texture regions and uneven feature distribution on system accuracy. For tracking both point and line features, an improved optical flow method was utilized. The initial value for the optical flow calculation was provided by the pose information from the inertial measurement unit (IMU) and the known depth of feature points, which replaces the traditional image pyramid iteration process, thereby saving computational resources and meeting the real-time requirements of the system.Finally, the proposed system was experimentally compared with leading open-source solutions in real-world scenarios to validate its real-time performance and accuracy. The experimental results demonstrate that the proposed algorithm can provide a highly robust localization solution for robots in applications such as industrial inspection and warehouse logistics, indicating significant potential for industrial application.
GONG Di-chen, JIANG Yang, ZHAO Jun-jie, HAO Xing-an.
Improved SLAM Algorithm for Feature Extraction and Matching of Points and Lines[J].
Science Technology and Engineering, 2026, 26(3): 1135-1146 DOI:10.12404/j.issn.1671-1815.2500427
即时定位与地图构建(simultaneous localization and mapping, SLAM)问题最早于1986年提出,其求解本质上是一个状态估计问题[1]。尽管视觉SLAM技术发展迅速,许多算法的实时性和精度满足生产要求,但是弱纹理场景中基于特征点法的视觉SLAM仍存在缺陷,过少的特征点无法满足位姿估计的要求[2]。Pumarola等[3]提出点线融合SLAM(point-line SLAM, PL-SLAM)系统,该方案在,ORB-SLAM(oriented fast and rotated brief simultaneous localization and mapping)框架中的一些模块中增加了线特征,构建了一套单目SLAM系统。PL-SLAM能在弱纹理环境中运行,甚至可以在连续三帧中仅使用线特征进行初始化,但该方案并未将线特征融入回环检测模块。He等[4]在视觉惯性导航系统(visual-inertial navigation system, VINS-Mono)的基础上融入了线特征,提出点线特征融合视觉惯性里程计(point and line feature fusion visual-inertial odometry, PL-VIO),该方案使用最常见的LSD ( line segment detector )线特征提取算法[5]和LBD ( line band descriptor )描述子[6]匹配,还提出了一种线三角化方法,相比原算法性能有所提升。PL-VIO只使用特征点进行初始化,线特征只用于后端,且没有构建回环模块,不能消除累计误差。Gomez-Ojede等[7]提出一种结合点线特征的双目PL-SLAM方案,这是一个完整的PL-SLAM框架,利用长度和方向剔除差异较大的匹配,得到最终匹配结果。Fu等[8]提出了PL-VINS方法,该方法调整了LSD算法的内部参数,并且认为长度更长的线特征更可靠,剔除短线特征,修改后的算法提取速度明显提升,使得SLAM系统的实时性得以保证,且定位误差比VINS方法更小。
为此,现提出一种基于VINS-Mono框架[9]下的线特征提取算法EHPL-VINS,该算法对线特征提取算法(efficient line segment detector and descriptor,ELSED)进行增加短线合并、梯度阈值参数调整,线特征均匀化剔除策略,能提高算法在低纹理环境下的适应性,并改善长线被分割成短线的现象;还提出融合IMU信息的改进型线特征光流法,可提高特征跟踪的速度和鲁棒性。
为了测试改进后的线特征提取算法在完整SLAM系统中的性能,选取EuRoC数据集中Machine Hall 01(MH01)、Machine Hall 04(MH04)、Machine Hall 05(MH05)、Vicon Room 1-01(V1-01)、Vicon Room 1-03(V1-03)、Vicon Room 2-03(V2-03)4个序列进行实验。其中Machine Hall 01和Vicon Room 1-01场景简单,可用于测试算法在普通环境中的性能,其他4个序列运动模糊、纹理差且场景黑暗,符合弱纹理场景的特征,用于测试算法在弱纹理场景中的运行效果。将采用改进后的线特征提取算法的SLAM系统命名为EMPL-VINS(enhanced multi point line feature visual inertial navigation system),与同为视觉惯性点线SLAM方案且效果较好的开源方案PL-VINS进行对比。
由表1可知本文改进的线特征提取算法的速度整体优于PL-VINS使用的LSD算法,经计算,EMPL-VINS平均速度比PL-VINS约快22%。其中在困难场景中两者差距比较明显,如序列Machine Hall 05中EMPL-VINS的速度提升达到43%;在简单场景中差距较小,原因是简单场景中的线特征数量较多,导致改进的线特征提取算法在线段合并和筛选步骤耗时较长,因而整体的提取速度差距不大。
进一步分析数据序列内每一帧提取耗时和线特征数目的分布情况。如图6所示为两种算法运行Machine Hall 05序列时每帧耗时分布。整体上EMPL-VINS速度更快,尤其在第500~800帧时EMPL-VINS的速度提升最明显。这时数据集大致运行至厂房外,图像中可提取的线特征相对较少,由前可知在线特征合并和筛选步骤的耗时较少,所以与PL-VINS耗时差距更大。
需要测试改进型光流法跟踪线特征的效果,验证其可行性。因此在采用传统LK光流法的EMPL-VINS系统的基础上,改用本文提出的改进型光流法ELPL-VINS(enhanced line feature pyramid LK optical flow visual inertial navigation system)。运行数据集序列Machine Hall 03(MH03)和Machine Hall 05(MH05),分别从这两个序列中各选取一帧,统计该帧图像中的线特征和此后连续10帧的跟踪情况,统计结果中不包括后续帧新提取的线特征如图8所示,其中绿线表示当前帧中提取的线特征,红线表示历史帧中的线特征。
如表3统计了跟踪过程中每帧能跟踪到的线特征数量,Machine Hall 03(MH03)序列第一帧中提取了412个线特征,在第二帧中追踪到了398个线特征,经过10帧后仍能跟踪到229个线特征;Machine Hall 05(MH05)序列第一帧中提取了215个线特征,第二帧中追踪到了201个线特征,经过10帧后仍跟踪到109个线特征。本文改进型光流法基本能成功跟踪到90%以上的线特征,跟踪效果较好,可以应用于后续的位姿估计和优化。
如图13所示为各个系统运行Machine Hall 05序列时的位置和姿态轨迹图,可知得各个系统的运行结果与真值的重合度都比较高。
对各个SLAM系统的实时性能进行对比,如表8所示为各个SLAM系统运行Machine Hall 05序列时各模块耗时。在线特征检测和跟踪模块,EHPL-VINS算法速度最快,之后是EPLF-VINS;本地VIO模块稍快于其他算法;回环检测模块中基于点线特征的系统都慢于VINS-Mono,原因是点线SLAM的回环模块包含线特征的回环检测,增加了计算量。同样运行其他五个序列并统计分析,可以得出相比其他基于点线特征的SLAM系统,本文中EHPL-VINS系统实时性最好。
LiHui, WangJingang, ZhangXiaojun. An improved SLAM algorithm for smart cars under weak texture in underground garages[J]. Science Technology and Engineering, 2021, 21(24): 10369-10375.
[5]
PumarolaA, VakhitovA, AgudoA, et al. PL-SLAM:real-time monocular visual SLAM with points and lines[C]// 2017 IEEE International Conference on Robotics and Automation (ICRA). New York: IEEE, 2017: 7989714.
[6]
HeY, ZhaoJ, GuoY, et al. PL-VIO: tightly-coupled monocular visual-inertial odometry using point and line features[J]. Sensors, 2018, 18(4): 1159.
[7]
vonGioi R G, JakubowiczJ, MorelJ, et al. LSD: a line segment detector[J]. Image Processing on Line, 2012, 2: 35-55.
[8]
ZhangL, KochR. An efficient and robust line segment matching approach based on LBD descriptor and pairwise geometric consistency[J]. Journal of Visual Communication and Image Representation, 2013, 24(7): 794-805.
[9]
Gomez-OjedaR, MorenoF, Zuniga-NoelD, et al. PL-SLAM: a stereo SLAM system through the combination of points and line segments[J]. IEEE Transactions on Robotics, 2019, 35(3): 734-746.
[10]
FuQ, WangJ, YuH, et al. PL-VINS: real-time monocular visual-inertial SLAM with point and line features[EB/OL]. arXiv Preprint, arXiv:2009. 07462, 2020.
[11]
QinT, LiP, ShenS. Vins-mono: a robust and versatile monocular visual-inertial state estimator[J]. IEEE Transactions on Robotics, 2018, 34(4): 1004-1020.
JiangLin, XiaXuhong, HanLu, et al. A stereo simultaneous localization and mapping method based on point-line feature fusion[J]. Science Technology and Engineering, 2020, 20(12): 4787-4792.
[14]
SuárezI, BuenaposadaJ M, BaumelaL. ELSED: enhanced line segment drawing[J]. Pattern Recognition, 2022, 127: 108619.
LinLimong, WangMei. A binocular vision SLAM algorithm with improved dot and line features[J]. Computer Measurement and Control, 2019, 27(9): 156-162.
[17]
LucasB D, KanadeT. An iterative image registration technique with an application to stereo vision[C]// IJCAI’81: 7th International Joint Conference on Artificial Intelligence. New York: IEEE, 1981: 674-679.
[18]
BouguetJ Y. Pyramidal implementation of the affine lucas kanade feature tracker description of the algorithm[J]. Intel Corporation, 2001, 5(4): 1-10.
[19]
XuL, YinH, ShiT, et al. EPLF-VINS: real-time monocular visual-inertial SLAM with efficient point-line flow features[J]. IEEE Robotics and Automation Letters, 2023, 8(2): 752-759.