近年来煤矿产业逐渐趋于自动化与智能化，视觉SLAM作为机器人领域的重要支撑，已被广泛用于煤矿井下开展各项工作。闭环检测作为视觉SLAM的关键组成部分，它通过摄像机采集周围环境信息来矫正自身位姿，帮助机器人在煤矿井下构建全局一致性的环境地图。而目前现有的闭环检测研究方法存在鲁棒性差、准确率低且耗时长等问题，为满足井下闭环检测准确性和实时性的要求，本文进行了如下研究。

（1）针对传统的闭环检测算法存在准确率低、鲁棒性差的问题，提出了基于卷积神经网络的闭环检测算法。即使用Faster RCNN网络代替传统手工设计特征的方式来提取煤矿数据集的图像特征，通过对比各卷积层提取图像特征的性能，选择在各数据集上表征能力较强的conv3层作为图像特征提取器，从而提高闭环检测的准确性和鲁棒性。

（2）针对网络提取的图像特征存在局部特征信息丢失的问题，提出了基于Faster RCNN-ROIs的闭环检测算法。即使用RPN网络结合增强注意力机制对网络提取的图像特征进行聚类融合，生成特征图的局部感兴趣区域。通过提取图像中的重要信息，从而使闭环检测的准确率得到进一步提高，但无法满足实时性的要求。

（3）针对现有的闭环检测算法在进行图像特征提取与匹配过程中耗时过长问题，提出了基于Faster RCNN-ROIs-LSH的闭环检测算法。即对感兴趣区域的图像特征构建哈希函数，利用局部敏感哈希算法对高维图像特征进行降维并构建哈希表，在保证高准确率的同时实现了对高维特征的降维。实验表明，经过降维处理后，本文算法的实时性提高了29.27%。最后，将本文算法在自建的三组煤矿井下数据集上与其他算法进行对比实验，进一步证明了本文算法在准确率与实时性方面均优于其他算法。

综上所述，本文提出的基于Faster RCNN-ROIs-LSH的闭环检测算法在自建的煤矿井下数据集上表现优异，在一定程度上提高了煤矿井下SLAM闭环检测算法的准确性与实时性。

In recent years, the coal mining industry has gradually become more automated and intelligent. As an important support in the field of robotics, visual SLAM has been widely used to carry out various work in coal mines. As a key component of visual SLAM, loop closure detection uses cameras to collect information about the surrounding environment to correct its own posture and help robots build a globally consistent environment map in coal mines. However, the existing loop closure detection research methods have problems such as poor robustness, low accuracy and long time. In order to meet the requirements of accuracy and real-time of downhole loop closure detection, the following research is carried out in this paper.

Aiming at the problems of low accuracy and poor robustness of traditional loop closure detection algorithms, a loop closure detection algorithm based on convolutional neural network is proposed. That is, the Faster RCNN network is used to replace the traditional hand-designed features to extract the image features of the coal mine data set. By comparing the performance of each convolutional layer to extract image features, the conv3 layer with stronger representation ability on each data set is selected as the image feature extraction. This improves the accuracy and robustness of loop closure detection.

Aiming at the problem of the loss of local feature information in the image features extracted by the network, a loop closure detection algorithm based on Faster RCNN-ROIs is proposed. That is, using the RPN network combined with the enhanced attention mechanism to cluster and fuse the image features extracted by the network to generate the local area of interest of the feature map. By extracting important information in the image, the accuracy of loop closure detection is further improved, but it cannot meet the requirements of real-time performance.

Aiming at the problem that the existing loop closure detection algorithms take too long in the process of image feature extraction and matching, a loop closure detection algorithm based on Faster RCNN-ROIs-LSH is proposed. That is, a hash function is constructed for the image features of the region of interest, and the locality-sensitive hashing algorithm is used to reduce the dimension of high-dimensional image features and build a hash table, which realizes the dimension reduction of high-dimensional features while ensuring high accuracy. Experiments show that after dimensionality reduction, the real-time performance of the algorithm in this paper is improved by 29.27%. Finally, the algorithm in this paper is compared with other algorithms on three sets of self-built coal mine underground data sets, which further proves that the algorithm in this paper is superior to other algorithms in terms of accuracy and real-time performance.

To sum up, the loop closure detection algorithm based on Faster RCNN-ROIs-LSH proposed in this paper has excellent performance on the self-built coal mine underground data set, and to a certain extent improves the accuracy and real-time performance of the loop closure detection algorithm of SLAM underground coal mines.

[1]马晓燕. 煤矿井下巡检机器人的研究[J]. 煤炭技术, 2021, 40(10): 169-172.

[2]胡而已, 葛世荣. 煤矿机器人研发进展与趋势分析[J]. 智能矿山, 2021, 2(01): 59-74.

[3]王路明, 常振兴. 机器人技术在煤矿中的应用及发展趋势[J]. 煤炭技术, 2021, 40(04): 151-153.

[4]Hong Seonghun, Park SoonYong, Lee Sejin, et al. Special issue on recent advancements in simultaneous localization and mapping (SLAM) and its applications[J]. ETRI Journal, 2021, 43(4): 577-579.

[5]吴涛. 用于移动机器人的视觉SLAM综述[J]. 数据通信, 2022(01): 48-51.

[6]Yang Li, Chao Ping Chen, Yuan Liu, et al. 67-4: Visual Simultaneous Localization and Mapping with Deep Neural Network Based Loop Detection for Augmented Reality[J]. SID Symposium Digest of Technical Papers, 2020, 51(1): 1005-1008.

[7]余宇, 胡峰. 基于深度学习的视觉SLAM回环检测方法[J]. 计算机工程与设计, 2020(2): 530-535.

[8]Smith, R. C. Hamish, Cheeseman, et al. On the Representation and Estimation of Spatial Uncertainty[M]. Sage Publications, Inc. 1986, 5(4): 56-68.

[9]Johannsson H, Kaess M, Fallon M, et al. Temporally scalable visual SLAM using a reduced pose graph[C]//2013 IEEE International Conference on Robotics and Automation. IEEE, 2013: 54-61.

[10]Heng L, Lee G H, Pollefeys M. Self-calibration and visual SLAM with a multi-camera system on a micro aerial vehicle[J]. Autonomous Robots, 2014, 39(3): 259-277.

[11]Chan S H, Wu P T, Fu L C. Robust 2D indoor localization through laser SLAM and visual SLAM fusion[C]//2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2018: 1263-1268.

[12]Ding Z, Huang R, Hu B. Robust Indoor SLAM based on Pedestrian Recognition by Using RGB-D Camera[C]// 2019 Chinese Automation Congress (CAC). 2019: 103-109.

[13]Cui L, Ma C. SOF-SLAM: A semantic visual SLAM for dynamic environments[J]. IEEE access, 2019, 7: 166528-166539.

[14]Yang Y, Tang D, Wang D, et al. Multi-camera visual SLAM for off-road navigation[J]. Robotics and Autonomous Systems, 2020, 128: 1-10.

[15]Guan P, Cao Z, Chen E, et al. A real-time semantic visual SLAM approach with points and objects[J]. International Journal of Advanced Robotic Systems, 2020, 17(1): 1-10.

[16]Hu Xiao, Lang Jochen. DOE-SLAM: Dynamic Object Enhanced Visual SLAM[J]. Sensors, 2021, 21(9): 3091-3091.

[17]Liu Y, Miura J. RDMO-SLAM: Real-time visual SLAM for dynamic environments using semantic label prediction with optical flow[J]. IEEE Access, 2021, 9: 106981-106997.

[18]鲍振强, 李艾华, 崔智高, 等. 融合多层次卷积神经网络特征的闭环检测算法[J]. 激光与光电子学进展, 2018, 55(11): 375-381.

[19]Baeza-Yates R , Ribeiro-Neto B . Modern Information Retrieval: Addison Wesley[J]. Computer Science & Information Technology, 1999: 1-6.

[20]Bampis L, Amanatiadis A, Gasteratos A. High order visual words for structure-aware and viewpoint-invariant loop closure detection[C]//2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2017: 4268-4275.

[21]Lee S, Jo H G, Cho H M, et al. Robust Visual Loop Closure Detection with Repetitive Features[C]//2018 15th International Conference on Ubiquitous Robots (UR). IEEE, 2018: 891-895.

[22]Garcia-Fidalgo E, Ortiz A. ibow-lcd: An appearance-based loop-closure detection approach using incremental bags of binary words[J]. IEEE Robotics and Automation Letters, 2018, 3(4): 3051-3057.

[23]Lee S, Jo H G, Cho H M, et al. Visual Loop Closure Detection over Illumination Change[C]//2019 16th International Conference on Ubiquitous Robots (UR). IEEE, 2019: 77-80.

[24]M. Labbé, F. Michaud. RTAB‐Map as an open‐source lidar and visual simultaneous localization and mapping library for large‐scale and long‐term online operation[J]. Journal of Field Robotics, 2019, 36(2): 416-446

[25]Memon A R, Wang H, Hussain A. Loop closure detection using supervised and unsupervised deep neural networks for monocular SLAM systems[J]. Robotics and Autonomous Systems, 2020, 126: 4-27.

[26]李伟, 任孟瀚, 黄威豪, 等. 基于改进M-ORB的视觉SLAM直接-闭环检测算法[J]. 智能科学与技术学报, 2021, 3(4): 482-491.

[27]Sünderhauf N, Shirazi S, Dayoub F, et al. On the performance of convnet features for place recognition[C]//2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2015: 4297-4304.

[28]Xia Y, Li J, Qi L, et al. An evaluation of deep learning in loop closure detection for visual SLAM[C]//2017 IEEE international conference on internet of things (iThings) and IEEE green computing and communications (GreenCom) and IEEE cyber, physical and social computing (CPSCom) and IEEE smart data (SmartData). IEEE, 2017: 85-91.

[29]Zhang W, Liu G, Tian G. Hha-based cnn image features for indoor loop closure detection[C]//2017 Chinese Automation Congress (CAC). IEEE, 2017: 4634-4639.

[30]Han F, Wang H, Huang G, et al. Sequence-based sparse optimization methods for long-term loop closure detection in visual SLAM[J]. Autonomous Robots, 2018, 42(7): 1323-1335.

[31]Liu H, Zhao C, Huang W, et al. An end-to-end siamese convolutional neural network for loop closure detection in visual SLAM system[C]//2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018: 3121-3125.

[32]Xiang R, Liu Y, Zhang Q, et al. Spatial pyramid pooling based convolutional autoencoder network for loop closure detection[C]//2019 IEEE International Conference on Real-time Computing and Robotics (RCAR). IEEE, 2019: 714-719.

[33]王凯. 基于深度学习的视觉SLAM闭环检测研究[D]. 哈尔滨: 哈尔滨工程大学，2019.

[34]胡年宗, 伍世虔, 张亦明. 基于卷积神经网络的SLAM回环检测算法研究[J]. 计算机仿真, 2020, 37(05): 282-286.

[35]Khaliq Ahmad, Ehsan Shoaib, Chen Zetao, et al. A Holistic Visual Place Recognition Approach Using Lightweight CNNs for Significant ViewPoint and Appearance Changes[J]. IEEE Transactions on Robotics, 2020: 1-9.

[36]Xiong F, Ding Y, Yu M, et al. A Lightweight sequence-based Unsupervised Loop Closure Detection[C]//2021 International Joint Conference on Neural Networks (IJCNN). IEEE, 2021: 1-8.

[37]Kim J J Y, Urschler M, Riddle P J, et al. SymbioLCD: Ensemble-Based Loop Closure Detection using CNN-Extracted Objects and Visual Bag-of-Words[C]//2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021: 5425-5425.

[38]Zhao Yipu and Vela Patricio A.. Good Feature Matching: Toward Accurate, Robust VO/VSLAM With Low Latency[J]. IEEE Transactions on Robotics, 2020, : 1-19.

[39]汪丹, 石朝侠, 王燕清. 基于非监督深度学习的闭环检测方法[J]. 计算机科学, 2020, 47(10): 228-232.

[40]Khaliq A, Ehsan S, Chen Z, et al. A Holistic Visual Place Recognition Approach Using Lightweight CNNs for Significant View Point and Appearance Changes[J]. IEEE Transactions on Robotics, 2019, PP(99):1-9.

[41]Zhou B, Lapedriza A, Khosla A, et al. Places: A 10 Million Image Database for Scene Recognition[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2018: 1-1.

[42]呼香艳. 基于粒子群优化算法的 SLAM 闭环检测方法研究[D]. 西安: 西安电子科技大学, 2018.

[43]Vallve J, Sola J, Andrade-Cetto J. Graph SLAM sparsification with populated topologies using factor descent optimization[J]. IEEE Robotics and Automation Letters, 2018, 3(2):1322-1329.

[44]赵凯, 朱愿, 谢枫. 基于改进 RANSAC 的点云关键点匹配[J]. 智能计算机与应用，2018, 8(6): 18-21．

[45]Samiksha Choyal and Ajay Kumar Singh. Speech based Object Identification using Region Proposal Faster RCNN Algorithm[J]. International Journal of Recent Technology and Engineering (IJRTE), 2019, 7(6s): 943-946.

[46]Baifan Chen, Dian Yuan, Chunfa Liu, et al. Loop Closure Detection Based on Multi-Scale Deep Feature Fusion[J]. Applied Sciences, 2019, 9(6): 1120-1120.

[47]何元烈, 陈佳腾, 曾碧. 基于精简卷积神经网络的快速闭环检测方法[J]. 计算机工程, 2018, 44(06): 182-187.

[48]Tian Ying Zhong et al. Robust identification of weld seam based on region of interest operation[J]. Advances in Manufacturing, 2020, 8(4): 473-485.

[49]Dongcan Zhang, Guoliang Zhang, Junxue Li, et al. Influence of Depth and Structure of Convolutional Neural Network on Loop Closure Detection[J]. World Scientific Research Journal,2021,7(6): 463-472.

[50]Peng W A, Jw A, Chen W A, et al. A novel fusing semantic- and appearance-based descriptors for visual loop closure detection[J]. Optik, 2021: 1-8

[51]Zheng Bolong et al. PM-LSH:A fast and accurate LSH framework for high-dimensional approximate NN search[J]. Proceedings of the VLDB Endowment, 2020, 13(5) : 643-655.

[52]Jiaohua Qin et al. An Encrypted Image Retrieval Method Based on Harris Corner Optimization and LSH in Cloud Computing.[J]. IEEE Access, 2019, 7: 24626-24633.

[53]Mehmet Ali Abdulhayoglu and Bart Thijs. Use of locality sensitive hashing (LSH) algorithm to match Web of Science and Scopus.[J]. Scientometrics, 2018, 116(2): 1229-1245.