School of Control Engineering,Northeastern University at Qinhuangdao,Qinhuangdao 066004,China
Show less
文章历史+
Received
Published
2025-02-02
2026-05-20
Issue Date
2026-09-09
PDF (3574K)
摘要
针对传统视觉simultaneous localization and mapping(SLAM)回环检测算法在光照变化、动态场景及视角变化等复杂环境下容易出现定位精度下降和累积误差增大的问题,提出一种基于MobileNetV3的回环检测算法。利用预训练的MobileNetV3模型提取图像特征,并通过主成分分析(PCA)和白化处理降低特征向量维度,提升计算效率。采用余弦相似度计算图像特征之间的相似度矩阵,并根据设定阈值判断是否出现回环。在New College和City Centre数据集上的实验结果表明,基于MobileNetV3的回环检测算法在几种对比算法中表现最优,与基于视觉词袋模型(BOVW)的回环检测算法相比,所提算法在两个数据集上的检测准确率分别提高18.5%和19.3%,检测速度分别提高30.6%和34.4%,能更好满足视觉SLAM对准确性和实时性的要求。最后将此算法应用到oriented FAST and rotated brief SLAM2(ORB-SLAM2)中,替换其原有的基于视觉词袋模型的回环检测算法,并在EuRoC数据集上测试改进后的ORB-SLAM2算法,实验结果表明,改进后的ORB-SLAM2算法定位精度比原算法提升23.8%,生成的轨迹曲线更接近真实轨迹,验证了所提算法在SLAM系统中的可行性和有效性。
Abstract
This work proposes algorithm based on MobileNetV3 to address the challenges of decreasing localization accuracy and increasing cumulative errors in traditional visual simultaneous localization and mapping (SLAM) loop closure detection algorithms in complex environments such as illumination variations, dynamic scenes, and viewpoint changes. A pretrained MobileNetV3 model is utilized to extract robust image features, followed by principal component analysis (PCA) and whitening to reduce the dimensionality of feature vectors and improve computational efficiency. Cosine similarity is employed to compute a similarity matrix of image features, and loop closures are identified based on predefined thresholds. Experimental results demonstrated that our MobileNetV3-based algorithm outperformed several comparative methods. Compared with the Bag-of-Visual-Words (BOVW) based approach, when using the New College and City Centre datasets our method achieved improvements in detection accuracy of 18.5 %and 19.3%, and enhanced detection speed of 30.6% and 34.4%, respectively. These results meet the accuracy and real-time performance requirements of visual SLAM systems. Furthermore, the algorithm was integrated into ORB-SLAM2 by replacing its original BOVW loop closure module. Evaluations using the EuRoC dataset showed that the enhanced ORB-SLAM2 achieved a 23.8% improvement in localization accuracy, with the estimated trajectories significantly closer to the Ground Truth. These results validate the feasibility and effectiveness of our algorithm for use within SLAM systems.
在自主机器人系统中,同步定位与建图(simultaneous localization and mapping,SLAM)技术在近几十年来一直是研究的重点[1]。SLAM技术主要解决定位和环境地图的重建两个问题[2]。根据传感器不同SLAM技术可分为激光SLAM和视觉SLAM,其中激光SLAM依赖于激光雷达传感器,而视觉SLAM是以相机作为传感器[3]。经典视觉SLAM流程包括4个主要任务[4]:前端视觉里程计、后端(非线性)优化、回环检测、建图。其中回环检测(loop closure detection,LCD)是其中一个关键环节,它通过识别之前访问过的地点来减少累积误差,从而提高SLAM系统的定位精度与鲁棒性[5]。
回环检测有两种常用的方法,一种是基于视觉里程计的方法,另一种是基于外观的方法。基于外观的方法是一种常用的回环检测技术,因为与基于视觉里程计的方法相比,其可以减少误差的累积。它支持对后端位姿优化施加约束,并在建立精确地图中起着关键作用,使回环检测成为一个独立模块[6]。大部分现有的回环检测方法都利用视觉词袋模型(bag of visual words,BOVW)进行位置识别,比如oriented FAST and rotated brief SLAM (ORB-SLAM) [7],visual⁃inertial navigation system⁃mono(VINS⁃Mono)[8]和Kimera[9]。该方法的过程主要分为3个步骤:首先用算法(scale⁃invariant feature transform(SIFT)[10], speed⁃up robust features(SURF)[11], oriented fast and rotated brief(ORB)[12])从不同类别的图像中提取特征向量,这些向量代表图像中的局部不变特征点;然后利用K⁃means聚类算法将具有相似词义的视觉词汇组合起来,构建包含K个单词的单词列表;最后统计单词列表中每个单词在图像中出现的次数,从而用K维的数值向量来表示图像。然而,传统的视觉回环检测方法依赖手工制作的特征,大多数基于手工制作特征的闭环检测方法主要在静态环境中有效,难以处理视点的剧烈变化和照明条件的变化,这使得它们在复杂动态环境中识别回环的准确性较低[13]。
完成PCA降维和白化处理后,下一步是计算相似度矩阵。常用的相似度度量方法有欧氏距离和余弦距离。相较于对向量模长敏感的欧氏距离,余弦距离通过衡量向量间的夹角来计算其方向上的相似性,更适合用于高维特征空间的相似性评估。因此,本文选用余弦距离来度量两帧图像特征之间的相似程度。假设在n维空间中存在两个非零向量 x 和 y,则 x 和 y 的余弦距离公式如下。
PANZ H, HOUJ Y, YUL. Optimization RGB⁃D 3⁃D reconstruction algorithm based on dynamic SLAM[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-13.
[2]
FUP, LIS X, YUT Y, et al. An initialization method for monocular visual⁃inertial odometry based on uncertainty state[C]//2022 4th International Conference on Communications, Information System and Computer Engineering. Shenzhen: IEEE, 2022: 522-528.
[3]
TANGM, CHENZ, YINF L. Robot tracking in SLAM with Masreliez⁃Martin unscented Kalman filter[J]. International Journal of Control, Automation and Systems, 2020, 18(9): 2315-2325.
[4]
QIH X, WANGC H, LIJ W, et al. Loop closure detection with CNN in RGB⁃D SLAM for intelligent agricultural equipment[J]. Agriculture, 2024, 14(6): 949.
[5]
ZHANGX W, WANGL, SUY. Visual place recognition: a survey from deep learning perspective[J]. Pattern Recognition, 2021, 113: 107760.
[6]
ZHOUY H, SUNM L. A visual SLAM loop closure detection method based on lightweight siamese capsule network[J]. Scientific Reports, 2025, 15: 7644.
[7]
MUR⁃ARTALR, MONTIELJ M M, TARDOSJ D. ORB⁃SLAM: a versatile and accurate monocular SLAM system[J]. IEEE Transactions on Robotics, 2015, 31(5): 1147-1163.
[8]
QINT, LIP L, SHENS J. VINS⁃Mono: a robust and versatile monocular visual⁃inertial state estimator[J]. IEEE Transactions on Robotics, 2018, 34(4): 1004-1020.
[9]
ROSINOLA, ABATEM, CHANGY, et al. Kimera: an open⁃source library for real⁃time metric⁃semantic localization and mapping[C]//2020 IEEE International Conference on Robotics and Automation (ICRA). Paris: IEEE, 2020: 1689-1696.
[10]
LOWED G. Distinctive image features from scale⁃invariant keypoints[J]. International Journal of Computer Vision, 2004, 60: 91-110.
[11]
BAY H, ESS A, TUYTELAARST, et al. Speeded⁃up robust features (SURF)[J]. Computer Vision and Image Understanding, 2008, 110(3): 346-359.
[12]
RUBLEEE, RABAUDV, KONOLIGEK, et al. ORB: an efficient alternative to SIFT or SURF[C]//2011 International Conference on Computer Vision. Barcelona: IEEE, 2011: 2564-2571.
[13]
JIANGS H, ZHOUZ K, SUNS J. Loop closure detection based on compressed ConvNet features in dynamic environments[J]. Applied Sciences, 2023, 14: 8.
[14]
WANGX, CHRISTIEM, MARCHANDE. Binary graph descriptor for robust relocalization on heterogeneous data[J]. IEEE Robotics and Automation Letters, 2022, 7(2): 2008-2015.
[15]
MAJ Y, WANGS Y, ZHANGK N, et al. Fast and robust loop⁃closure detection via convolutional auto⁃encoder and motion consensus[J]. IEEE Transactions on Industrial Informatics, 2022, 18(6): 3681-3691.
[16]
ZHANGX, ZHENGL M, TANZ H, et al. Loop closure detection based on residual network and capsule network for mobile robot[J]. Sensors, 2022, 22: 7137.
[17]
DIANS Y, YINY J, WUC, et al. Loop closure detection based on local⁃global similarity measurement strategies[J]. Journal of Electronic Imaging, 2022, 31(2): 023004.
YANGJ, HUW, ZHUW F. Visual SLAM loop closure detection algorithm based on improved MobileNetV3[J]. Computer and Modernization, 2024(10): 21-26. (in Chinese)
[20]
HOWARDA, SANDLERM, CHUG, et al. Searching for MobileNetV3[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 1314-1324.
[21]
SANDLERM, HOWARDA, ZHUM L, et al. MobileNetV2: inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 4510-4520.
[22]
CAMPOSC, ELVIRAR, GOMEZ RODRÍGUEZJ J, et al. ORB⁃SLAM3: an accurate open⁃source library for visual, visual⁃inertial, and multimap SLAM[J]. IEEE Transactions on Robotics, 2021, 37(6): 1874-1890.