矿井监控视频分析作为智慧矿山的重要研究内容，人员检测与统计是其中的重要环节。由于煤矿井下存在复杂的光照条件和狭长的巷道，监控图像容易出现照度低、光照不均、小目标与相互遮挡等情况，致使现有检测及计数算法应用于矿井图像时，存在目标识别困难，ID转换频繁等问题。因此，开展基于矿井图像的人员检测和跟踪计数算法研究对于煤矿安全生产管理具有重要意义。本文主要工作如下：

针对矿井图像容易出现照度低、光照不均、光晕伪影等情况，在图像数据预处理单元中构建自适应增强模块不断优化待检测图像的亮度、对比度及边缘清晰度。自适应增强模块通过一个弱监督的CNN层预测图像调节所需的参数，并通过检测端的损失反向传播进行不断的学习与调节，使得输入图像的亮度、色调等特征向有利于检测的方向调节。实验结果表明，本文增强算法能够自适应改善待检测图像的质量，有效去除图像中的大部分噪声，提升检测算法对于矿井复杂环境的适应能力。

针对矿井人员出现遮挡和小目标情况致使检测效果不佳的问题，提出改进YOLOv7的矿井人员检测算法。首先提出基于通道重组与特征关注的复杂场景检测方式，构建STR_shuffleNet模块促进组间信息流动的同时增强算法的全局感受野，提升算法的抗遮挡性能；其次提出基于卷积与自注意力混合模块的小目标检测方式，引入AC_BiFPN结构兼顾全局与局部特征的同时提升不同尺度间的特征融合能力，提升算法的小目标检测性能；最后采用Efficient IOU Loss作为损失函数，解决了难易样本均衡与纵横比模糊定义的问题，提升算法速度的同时实现更加精准的定位。实验结果表明，本文检测算法较YOLOv7相比精度提升了3.7%，达到90.1%；速度提升了15.5%，达到68.4FPS。

针对矿井人员出现快速运动和尺度变化情况致使跟踪效果不佳的问题，提出改进DeepSORT的矿井人员跟踪计数模型。首先使用行人重识别网络OSNet强化外观模型的特征提取能力；然后采用指数移动平均策略实现外观状态的更新，提升匹配质量的同时减少时间的消耗；最后选用Complete IOU Loss优化检测结果与跟踪轨迹间的匹配，减少跟踪过程中的状态重新确认现象，进一步降低ID转换次数。实验结果表明，本文计数算法准确率达到88.4%，速度达到54.7FPS，可以对矿井人员实现实时准确的计数。

本文方法对矿井图像增强和人员检测计数有着一定的参考价值，对加快建设一批智能、安全、高效的现代化矿井起到积极推动作用。

Mine surveillance video analysis is an important research content of smart mine, and personnel detection and statistics are important links. Due to the complex illumination conditions and narrow and long roadway in coal mine, monitoring images are prone to low illumination, uneven illumination, small targets and mutual occlusion, etc. resulting in difficulties in target recognition and frequent ID conversion when the existing detection and counting algorithms are applied to mine images. Therefore, the research of personnel detection and tracking counting algorithm based on mine images is of great significance for mine safety production management. The main work of this thesis is as follows:

(1) In view of the low illumination, uneven illumination and halo artifacts in mine images, an adaptive enhancement module is constructed in the image data preprocessing unit to continuously optimize the brightness, contrast and edge clarity of the images to be detected. The adaptive enhancement module predicts the parameters required for image adjustment through a weakly supervised CNN layer, and learns and adjusts continuously through the loss backpropagation at the detection end, so that the brightness, tone and other features of the input image are adjusted in a direction conducive to detection. The experimental results show that the enhanced algorithm in this thesis can adaptively improve the quality of the image to be detected, effectively remove most of the noise in the image, and improve the adaptability of the detection algorithm to the complex mine environment.

(2) Aiming at the problem of poor detection effect caused by the occlusion of mine personnel and small targets, an improved mine personnel detection algorithm of YOLOv7 is proposed. Firstly, a complex scene detection method based on channel recombination and feature concern is proposed, and the STR_shuffleNet module is constructed to promote the information flow between groups and enhance the global sensitivity field of the algorithm to improve the anti-occlusion performance of the algorithm. Secondly, a small target detection method based on convolutional and self-attention hybrid module is proposed, and the AC_BiFPN structure is introduced to improve the feature fusion ability between different scales while taking into account the global and local features, and improve the small target detection performance of the algorithm. Finally, Efficient IOU Loss is used as a loss function to solve the problems of difficult and easy sample balancing and fuzzy definition of aspect ratio, improving the algorithm speed and achieving more accurate positioning. Experimental results show that compared with YOLOv7, the accuracy of the proposed detection algorithm is improved by 3.7%, reaching 90.1%. Speed increased by 15.5% to 68.4FPS.

(3) To solve the problem of poor tracking effect caused by rapid movement and scale change of mine personnel, an improved DeepSORT mine personnel tracking and counting model was proposed. Firstly, pedestrian re-recognition network OSNet was used to strengthen the feature extraction capability of the appearance model. Then the exponential moving average strategy is used to update the appearance status, improve the matching quality and reduce the time consumption. Finally, Complete IOU Loss is used to optimize the matching between the detection results and the tracking track, which reduces the status reconfirmation phenomenon in the tracking process and further reduces the ID conversion times. The experimental results show that the accuracy rate of this algorithm is 88.4% and the speed is 54.7FPS, which can realize the real-time and accurate counting of mine personnel.

The method in this thesis has a certain reference value for mine image enhancement and personnel detection and counting, and plays an active role in accelerating the construction of a batch of intelligent, safe and efficient modern mines.

[1]丁恩杰,俞啸,夏冰,等.矿山信息化发展及以数字孪生为核心的智慧矿山关键技术[J].煤炭学报,2022,47(1):564-578.

[2]张建中,郭军.智慧矿山工业互联网技术架构探讨[J].煤炭科学技术,2022,50(5):238-246.

[3]周晨晖.基于深度学习的煤矿复杂场景人员检测与统计分析方法研究[D].北京:中国矿业大学,2019.

[4]董锴文,孙彦景,陈岩,等.矿山生产作业场景视频结构化分析关键技术[J].煤炭学报,2021,46(11):3724-3735.

[5]杨克虎,龙启航,汪嘉文,等.基于自注意力机制的矿井次光照图像语义分割研究[J].矿业安全与环保,2023,50(5):9-18.

[6]De Y W,Ze X G. Enhancement algorithms for images in coal mines base on dark light denoising networks[C]//Proceedings of the 2019 Annual Meeting on Management Engineering(AMME),2019:1-7.

[7]王洪栋,郭伟东,朱美强,等.一种煤矿井下低照度图像增强算法[J].工矿自动化,2019,45(11):81-85.

[8]王满利,田子建.基于非下采样轮廓波变换的矿井图像增强算法[J].煤炭学报,2020,45(9):3351-3362.

[9]范伟强,刘毅.基于自适应小波变换的煤矿降质图像模糊增强算法[J].煤炭学报,2020,45(12):4248-4260．

[10]Kang C. Low-light images enhancement method in coal mine based on knowledge distillation[J]. Academic Journal of Engineering and Technology Science,2021,4(2):52-56.

[11]阮顺领,刘丹洋,白宝军,等.基于自适应MSRCP算法的煤矿井下图像增强方法[J].矿业研究与开发,2021,41(11):186-192.

[12]Li C,Liu J,Zhu J,et al. Mine image enhancement using adaptive bilateral gamma adjustment and double plateaus histogram equalization[J]. Multimedia Tools and Applications,2022,81(9):12643-12660.

[13]李藤,刘凯雷.基于深度学习的矿井下低照度图像增强算法研究[J].电子测试,2022,36(9):51-53+134.

[14]Tian F,Chen T,Zhang J. Research on Improved Retinex-Based Image Enhancement Method for Mine Monitoring[J]. Applied Sciences,2023,13(4):2672.

[15]王满利,张航,李佳悦,等.基于深度神经网络的煤矿井下低光照图像增强算法[J]. 煤炭科学技术,2023,51(9):231−241.

[16]朱代先,秋强,孔浩然,等.基于直线段检测和LT描述符的矿井图像线特征匹配算法[J/OL].工矿自动化:1-14[2024-02-29].

[17]王媛彬,郭亚茹,刘佳,等.基于注意力机制和空洞卷积的CycleGAN煤矿井下低照度图像增强算法[J/OL].煤炭科学技术:1-10[2024-1-23].

[18]王国法,庞义辉,任怀伟,等.智慧矿山系统工程及关键技术研究与实践[J/OL].煤炭学报:1-23[2024-02-02].

[19]Zhang K,Kang L,Chen X,et al. A review of intelligent unmanned mining current situation and development trend[J]. Energies,2022,15(2):513.

[20]Dong X,Feng S,Ze L, et al. A target detection model based on improved tiny-yolov3 under the environment of mining truck[J]. IEEE Access,2019,7: 123757-123764.

[21]李伟山,卫晨,王琳.改进的Faster RCNN煤矿井下行人检测算法[J].计算机工程与应用,2019,55(4):200-207.

[22]谭章禄,陈孝慈. RetinaNet图像识别技术在煤矿目标监测领域的应用研究[J].矿业安全与环保,2020,47(5):65-70．

[23]韩江洪,沈露露,卫星,等.基于轻量级CNN的井下视觉识别策略[J].合肥工业大学学报(自然科学版),2020,43(11):1469-1475+1562.

[24]Hao B N,Zhang L Y,Peng R. An improved vibe algorithm to detect personnel underground in coal mines[J]. Journal of Physics: Conference Series,2021,2025(1):012032.

[25]张小艳,郭海涛.基于改进混合高斯模型的井下目标检测算法[J].工矿自动化,2021,47(4):67-72.

[26]Li X Y,Wang S,Chen W,et al. Improved YOLOv4 network using infrared images for personnel detection in coal mines[J]. Journal of Electronic Imaging,2022,31(1):013017.

[27]Yuan H D,Li W. Machine learning-based multi-target detection method for the complex mine environment[J]. International Journal of Reasoning-based Intelligent Systems,2022, 14(4):201-207.

[28]寇发荣,肖伟,何海洋,等.基于改进YOLOv5的煤矿井下目标检测研究[J].电子与信息学报,2023,45(7):2642-2649.

[29]Sun C M,Li X Y,Chen J X,et al. Coal-Rock Image Recognition Method for Complex and Harsh Environment in Coal Mine Using Deep Learning Models[J]. IEEE Access,2023,11: 80794-80805.

[30]陈湘源,饶天荣,潘涛.基于半监督学习的煤矿井下行人检测模型[J].煤炭技术,2024,43(2):231-234.

[31]闫明伟.煤矿井下巷道行驶车辆车前障碍物检测方法研究[J].矿业安全与环保,2024,51(1):168-174.

[32]邵小强,李鑫,杨涛,等.改进YOLOv5s和DeepSORT的井下人员检测及跟踪算法[J].煤炭科学技术,2023,51(10):291-301.

[33]孙晓东.全矿井人工智能(AI)监管平台关键技术[J].煤矿安全,2023,54(12):220-226.

[34]张丽颖.矿井下目标尺度自适应跟踪算法研究[D].北京:中国矿业大学,2019.

[35]徐桂涛.煤矿视频监控系统中人员目标跟踪算法的研究与实现[D].北京:中国矿业大学,2019.

[36]孔乐天.井下矿工多目标检测与跟踪[D].太原:太原理工大学,2020.

[37]刘贝.矿井人员目标检测与跟踪研究[D].西安:西安科技大学,2020.

[38]Lu Y,Huang Q. Underground Coal Mine Target Tracking via Multi-Feature Joint Sparse Representation[J]. Journal of Computer and Communications,2021,9(3):118-132.

[39]陈伟,任鹏,田子建,等.基于注意力机制的无监督矿井人员跟踪[J].煤炭学报,2021,46(S1):601-608.

[40]周孟然,李学松,朱梓伟,等.井下矿工多目标检测与跟踪联合算法[J].工矿自动化,2022,48(10):40-47.

[41]张玉涛,张梦凡,史学强,等.基于深度学习的井下运动目标跟踪算法研究[J].煤炭工程,2022,54(10):151-155.

[42]Wang R,Li J,Xu Z. Dark-SORT:Multi-Person Tracking in Underground Coal Mines Using Adaptive Discrete Weighting[J]. IEEE Access,2023,11:139422-139438.

[43]You S Z,Zhu H,Li M G,et al. Tracking system for a coal mine drilling robot for low-illumination environments[J]. Applied Sciences,2022,13(1):568.

[44]范伟强,王雪瑾,张颖慧,等.改进YOLOv7和DeepSORT的井下人员检测与跟踪算法[J/OL].煤炭科学技术:1-13[2024-01-22].

[45]丁玲,缪小然,胡建峰,等.改进YOLOv8s与DeepSORT的矿工帽带检测及人员跟踪[J/OL].计算机工程与应用:1-11[2023-12-25].

[46]邵小强,杨涛,卫晋阳,等.改进同态滤波的矿井监控视频图像增强算法[J].西安科技大学学报,2022,42(6):1205-1213.

[47]程健,李昊,马昆,等.矿井视觉计算体系架构与关键技术[J].煤炭科学技术,2023,51(9):202-218.

[48]Xiang N,Nanxing W,Rumeng Z,et al. Tenglong Yuan blue and white texture extraction method based on adaptive gamma correction and K-means clustering segmentation coupled algorithm[J]. Journal of the Australian Ceramic Society,2023,60(1):1-11.

[49]Xiaoyan F,Zhanquan S,Engang T,et al. Medical image contrast enhancement based on improved sparrow search algorithm[J]. International Journal of Imaging Systems and Technology,2022,33(1):389-402.

[50]张云涛,刘惠萍,黄益铭,等.基于图像分割和色适应变换白平衡的水下图像增强[J].激光与光电子学进展,2023,60(14):37-43.

[51]杜新阳,张爱军.一种基于局部方差改进的自适应USM算法[J].电子设计工程,2023,31(21):191-195.

[52]顾文娟,丁灿,魏金,等.基于双边滤波MSR与AutoMSRCR融合的低光照图像增强[J].光学精密工程,2023,31(24):3606-3617.

[53]Yun J,Zeqi M,Chao W, et al. RSAP-Net: joint optic disc and cup segmentation with a residual spatial attention path module and MSRCR-PT pre-processing algorithm[J]. BMC Bioinformatics,2022,23(1):523-523.

[54]Zhang Z,Meng P,Yang Y, et al. Machine Learning-Augmented Micro-Defect Detection on Plastic Straw[J]. Micro,2023,3(2):484-495.

[55]孙林,陈圣,姚旭龙,等.矿井智能监控目标识别的图像增强方法与应用[J/OL].煤炭学报:1-12[2024-03-10].

[56]Shao X Q,Liu S B, Li X, et al. FM-YOLOv7: an improved detection method for mine personnel helmet[J]. Journal of Electronic Imaging,2023,32(3):033013.

[57]Ren S,He K,Girshick R,et al. Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks[C]//Proceedings of the IEEE Transactions on Pattren Analysis and Machine Intelligence.New York:IEEE,2017:1137-1149.

[58]Redmon J,Divvala S,Girshick R,et al. You only look once:Unified,real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.New York:IEEE,2016:779-788.

[59]Shao X Q,Liu S B,Li X,et al. Rep-YOLO: an efficient detection method for mine personnel[J]. Journal of Real-Time Image Processing,2024,21(2):1-16.

[60]Han J,Yang Y. L-Net: lightweight and fast object detector-based ShuffleNetV2[J]. Journal of Real-Time Image Processing,2021,18(6):1-12.

[61]Tiansheng C,Lingfei M. Swin-Fusion: Swin-Transformer with Feature Fusion for Human Action Recognition[J]. Neural Processing Letters,2023,55(8):11109-11130.

[62]Yuan Z,Youpeng S,Zheng W,et al. YOLOv7-RAR for Urban Vehicle Detection[J]. Sensors,2023,23(4):1801.

[63]Fuzhen Z,Yuying W,Jingyi C,et al. Target detection for remote sensing based on the enhanced YOLOv4 with improved BiFPN[J]. The Egyptian Journal of Remote Sensing and Space Sciences,2023,26(2):351-360.

[64]Fuzhen Z,Yuying W,Jingyi C,et al. Target detection for remote sensing based on the enhanced YOLOv4 with improved BiFPN[J]. The Egyptian Journal of Remote Sensing and Space Sciences,2023,26(2):351-360.

[65]何卫隆,王平,张爱华,等.基于目标检测的焊接电弧形态在线定量检测[J].浙江大学学报(工学版),2023,57(9):1903-1914.

[66]王昱,姚兴智,李斌,等.基于改进YOLO v7-tiny的甜椒畸形果识别算法[J].农业机械学报,2023,54(11):236-246.

[67]Zhiyong L,Xueqin J,Luyu S,et al.A Real-Time Detection Algorithm for Sweet Cherry Fruit Maturity Based on YOLOX in the Natural Environment[J]. Agronomy,2022,12(10):2482.

[68]Baek J,Hong S,Kim J,et al. Bayesian learning of a search region for pedestrian detection[J]. Multimedia Tools and Applications,2016,75(2):863-885.

[69]Yang W,Zhang X,Ma B,et al. An open dataset for intelligent recognition and classification of abnormal condition in longwall mining[J]. Scientific Data,2023,10(1):416.

[70]邵延华,张铎,楚红雨,等.基于深度学习的YOLO目标检测综述[J].电子与信息学报,2022,44(10):3697-3708.

[71]Liu W,Anguelov D,Erhan D,et al. SSD:single shot MultiBox detector[C]//Proceedings of European Conference on Computer Vision.Cham:Springer,2016:21-37.

[72]Luxuan B,Bo L,Jue W,et al. Multi-branch stacking remote sensing image target detection based on YOLOv5[J]. The Egyptian Journal of Remote Sensing and Space Sciences,2023,26(4):999-1008.

[73]Wannakamon P,Kittichai W,Arnon C,et al. Enhancing Caries Detection in Bitewing Radiographs Using YOLOv7[J]. Journal of digital imaging,2023,36(6):2635-2647.

[74]Quan J,Hui L. Silicon energy bulk material cargo ship detection and tracking method combining YOLOv5 and DeepSort[J]. Energy Reports,2023,9(S2):151-158.

[75]金立生,华强,郭柏苍,等.基于优化DeepSort的前方车辆多目标跟踪[J].浙江大学学报(工学版),2021,55(6):1056-1064.

[76]刘洋,闫冬梅,孟范伟.基于Transformer改进的两分支行人重识别算法[J].东北大学学报(自然科学版),2023,44(1):26-32.

[77]Cao X,Wang Z,Zheng B,et al. Improved UAV-to-Ground Multi-Target Tracking Algorithm Based on StrongSORT[J]. Sensors,2023,23(22):9239.

[78]Zhao Y,Li Y,Wang S. Asymmetric Deep Hashing for Person Re-Identifications[J]. Tsinghua Science and Technology,2022,27(2):396-411.

[79]霍旭,盖绍彦,洪濡,等.基于多阶段关联的多目标跟踪算法[J].仪器仪表学报,2023,44(11):205-214.

[80]Jiandong W,Yahui D,Shuangrui Z,et al. A High-Precision Vehicle Detection and Tracking Method Based on the Attention Mechanism[J]. Sensors,2023,23(2):724.

[81]高文印,文峰,单铭琦.基于改进CenterTrack的多目标跟踪算法[J].沈阳理工大学学报,2023,42(3):22-27.

[82]Zhongchao W,Xia Z,Yuxiang S,et al. Abnormal Behavior Monitoring Method of Larimichthys crocea in Recirculating Aquaculture System Based on Computer Vision[J]. Sensors,2023,23(5):2835.