行人重识别作为视频监控系统的关键技术，目前成为计算机视觉领域的研究热点之 一。行人重识别面临光照、视角、遮挡等问题，而视频行人重识别更是面临如何能够更 快、更准确的进行行人身份重识别的挑战。将深度学习应用到视频行人重识别的技术研究中能有效提升行人重识别速度和重识别的精度，具有一定的研究价值。

本文研究了一种洗煤厂内的跨镜头下多目标行人重识别算法。主要从洗煤厂内行人目标检测，目标跟踪，以及行人重识别三部分进行研究。主要解决的问题首先是行人重识别过程中出现的多个行人目标之间的遮挡导致的漏检问题；其次是跟踪过程中因需要对跟踪目标进行轨迹匹配更新，以及大量的参数计算，造成的跟踪速率缓慢的问题；最后是洗煤厂环境下行人重识别时因行人特征提取不足以及被遮挡造成的身份被识别为None 和误识别问题。

1）改进YOLOv5 目标检测算法。本文行人目标检测部分采用的为YOLOv5 算法，主要对YOLOv5 算法中目标框筛选NMS 算法进行改进。针对检测过程中，由于多个目标行人之间的互相遮挡造成的漏检问题，对原YOLOv5 算法中的NMS 算法进行替换，采用DIOU_NMS 算法作为检测目标框的筛选算法，使得漏检问题得到了改善。

2）改进DeepSORT 多目标跟踪算法。针对多目标行人在跟踪过程中由于轨迹匹配更新以及参数计算量大造成的跟踪延时问题，本文采用了 GhostNet 网络代替了原DeepSORT 特征提取网络，使得参数计算量由43.9MB 减少到了18.8MB，同时对注意力机制SENet 的激活函数进行改进，并把改进后的注意力机制模块融合到DeepSORT 算法中，进一步减少了参数计算复杂度。本部分主要是通过跟踪时间以及参数量的变化对实 验结果进行验证，最终的实验结果表明，改进后的参数计算量减少了57.2%，跟踪速率提升了18.01%。

3）改进FastReID 行人重识别算法。本文使用FastReID 算法对洗煤厂工人进行身份重识别，镜头下的行人在行走过程中，由于摄像机的角度以及周围工作设备的影响，会使得行人特征提取不足，从而造成身份被识别为None 以及误识别。

针对上述问题，本文首先对原行人图像库进行特征补充，截取镜头下的行人目标图像，进行相似度计算，保存低相似度图片，充实行人不同的特征；其次利用行人的前后帧像素坐标对误识别行人进行身份校正，降低行人识别为None 的次数和误识别率。提高行人重识别准确率。本部分主要通过对改进前后行人None 出现的次数，以及误识别次数，验证改进前后的效果，最终的实验结果表明，改进后行人身份被识别为None 的概率降低了15.8%，误识别率降低了40.3%。

通过对改进后的算法在洗煤厂环境下实验验证，本文研究的多摄像头多目标行人重识别算法能够实现行人身份的准确识别，在计算能力强的硬件环境下能够达到对洗煤厂环境下工人的实时跨摄像头跟踪、重识别的需求。

As a key technology in video surveillance systems, pedestrian recognition has become one of the research hotspots in the field of computer vision. Pedestrian re recognition faces issues such asillumination, viewing angle, and occlusion, while video pedestrian re recognition isfaced with the challenge of how to quickly and accurately re identify pedestrian identities. Applying deep learning to the research of video pedestrian re recognition technology can effectively improve the speed and accuracy of pedestrian re recognition, and has certain research value.

This article studies a cross camera multi target pedestrian re recognition  algorithm in a coal washing plant. The research mainly focuses on three parts: pedestrian target detection, target tracking, and pedestrian re recognition in the coal washing plant.The main problem to be solved is firstly the problem of missed detection caused by occlusion between multiple pedestrian argets during pedestrian re recognition; Secondly, during the tracking process, there is a problem ofslow tracking speed caused by the need for trajectory matching updates and a large amount of parameter calculations for the tracked target; Finally, there are issues with identification as None and misidentification caused by insufficient pedestrian feature extraction and occlusion during pedestrian recognition in the coalwashing plant environment.

1) ImproveYOLOv5 object detection algorithm.The pedestrian target detection section of this article adoptsthe YOLOv5 algorithm, mainly improving the NMS algorithm for target box filtering in the YOLOv5 algorithm. In response to the missed detection problem caused by mutual occlusion between multiple target pedestrians during the detection process, the NMS algorithm in the original YOLOv5 algorithm is replaced with DIOU_NMS algorithm, as a screening algorithmfor detecting target boxes, improvesthe problemofmissed detections.

2) Improved DeepSORT multi target tracking algorithm. In order to solve the tracking delay problem caused by the track matching update and the large amount of parametercalculation in the tracking process of multi-target pedestrians, this paper uses GhostNet network instead of the original DeepSORT feature extraction network, which reduces the amount of parameter calculation from 43.9MB to 18.8MB.At the same time, the activation function of the attention mechanism SENet is improved, and the improved attention mechanism module is integrated into the DeepSORT algorithm, Further reducing the complexity of parameter calculation. Thissection mainly verifiesthe experimental results by tracking the changesin time and parameter quantities. The final experimental results show that the improved parameter calculation reduces by 57.2%and the tracking rate increases by 18.01%.

3）Improved FastReID pedestrian recognition algorithm.This article uses the FastReID algorithm to re identify the identity of coal washing plant workers. During the walking process of pedestrians under the camera, due to the influence of the camera angle and surrounding work equipment,This can lead to insufficient pedestrian feature extraction, resulting in identity being recognized as None and misidentification. To address the above issues, this article first supplements the original pedestrian image library with features, captures pedestrian target images under the camera, performs similarity calculation, saves low similarity images, and enriches the different features of pedestrians; Secondly, the pixel coordinates of the front and back frames of pedestrians are used to correct the identity of misidentified pedestrians, reducing the number of times pedestrians are recognized as None and the misidentification rate. Improve the accuracy of pedestrian recognition. This section mainly verifies the effectiveness of the improvement by comparing the number of occurrences of pedestrian None before and after the improvement, as well as the number of false positives. The final experimental results show that the probability of pedestrian identity being recognized as None after the improvement isreduced by 15.8%, and the false positives are reduced by 40.3%.

Through experimental verification of the improved algorithm in a coal washing plant environment, the multi camera andmulti-objective pedestrian re recognition algorithmstudied in this paper can achieve accurate recognition of pedestrian identity, and can meet the real-time cross camera head tracking and re recognition requirements for workers in the coal washing plant environment in a hardware environmentwith strong computing power.

[1] 刘存信. 新时代 新征程 再扬帆——纪念安防行业蓬勃发展 40 周年[J]. 中

国安防, 2019(03):1.

[2] Gheissari N, Sebastian T B, Hartley R. Person reidentification using spatiotemporalappearance[C]//2006 IEEE Computer Society Conference on Computer Vision and PatternRecognition (CVPR'06).IEEE, 2006, 2: 1528-1535.

[3] 赵才荣, 齐鼎, 窦曙光, 等. 智能视频监控关键技术:行人再识别研究综述[J]. 中国科学:信息科学,2021,51(12):1979-2015.

[4] 杨锋,许玉,尹梦晓,符嘉成,等.基于深度学习的行人重识别综述[J].计算机应

用,2020,40(05):1243-1252.

[5] 叶 钰 , 王 正 , 梁 超 , 等 . 多 源 数 据 行 人 重 识 别 研 究 综 述 [J]. 自 动 化 学

报,2020,46(09):1869-1884.

[6] Li W, Zhu X, Gong S. Harmonious attention network for person re-identification[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018:2285-2294.

[7] Ye M, Shen J, Lin G, et al. Deep learning for person re-identification: A survey and outlook[J]. IEEE transactions on pattern analysis and machine intelligence, 2021, 44(6):

2872-2893.

[8] Wu D, Zheng S J, Zhang X P, et al. Deep learning-based methods for person re-identification:Acomprehensive review[J].Neurocomputing, 2019, 337: 354-371.

[9] Lee K, Sato D,Asakawa S, et al. Pedestrian detection with wearable camerasfor the blind:Atwo-way perspective[C]//Proceedings of the 2020 CHI Conference on Human FactorsinComputing Systems. 2020: 1-12.

[10]Chen L, Ma N, Wang P, et al. Survey of pedestrian action recognition techniques forautonomous driving[J].Tsinghua Science andTechnology, 2020, 25(4): 458-470.

[11]An F P, Liu J, Bai L. Pedestrian reidentification algorithm based on deconvolution network feature extraction-multilayer attention mechanism convolutional neural network[J]. Journalof Sensors, 2021, 2021: 1-12.

[12]Yuan L, Qiu Z. Mask-RCNN with spatial attention for pedestrian segmentation incyberphysicalsystems[J].ComputerCommunications, 2021, 180: 109-114.

[13]Zahid M, Khan M A,Azam F, et al. Pedestrian identification using motion-controlled deepneural network in real-time visualsurveillance[J]. SoftComputing, 2021: 1-17.

[14]Chen L, Ma N, Wang P, et al. Survey of pedestrian action recognition techniques forautonomous driving[J].Tsinghua Science andTechnology, 2020, 25(4): 458-470..

[15]钱亚萍,王凤随,熊磊.基于局部细化多分支与全局特征共享的无监督行人重识别方法

[J/OL].电子测量与仪器学报:1-10[2023-03-21].

[16]张勃兴,马敬奇,张寿明,等.利用全局与局部关联特征的行人重识别方法[J].电子测量与仪器学报,2022,36(06):205-212.

[17]李佳宾,李学伟,刘宏哲,等.基于局部特征关联与全局注意力机制的行人重识别[J].计算机工程,2022,48(01):245-252.

[18]张涛,易争明,李璇,等.一种基于全局特征的行人重识别改进算法[J].激光与光电子学进展,2020,57(24):324-330.

[19]SpeckerA, SchumannA,BeyererJ.An evaluation of design choicesfor pedestrian attributerecognition in video[C]//2020 IEEE International Conference on Image Processing (ICIP). IEEE, 2020: 2331-2335.

[20]Zeng H, Ai H, Zhuang Z, et al. Multi-task learning via co-attentive sharing for pedestrian attribute recognition[C]//2020 IEEE International Conference on Multimedia and Expo(ICME).IEEE, 2020: 1-6.

[21]Yang D, Zhang H, Yurtsever E, et al. Predicting pedestrian crossing intention with feature fusion and spatio-temporal attention[J]. IEEE Transactions on Intelligent Vehicles, 2022, 7(2): 221-230.

[22]Zhang J, Ren P, Li J. Deep template matching for pedestrian attribute recognition with the auxiliary supervision of attribute-wise keypoints[J]. arXiv preprint arXiv:2011.06798, 2020.

[23]Zhang L, Li K, Qi Y, et al. Local feature extracted by the improved bag of features method for person re-identification[J].Neurocomputing, 2021, 458: 690-700.

[24]Tang C, Sheng L, Zhang Z, et al. Improving pedestrian attribute recognition with weakly-supervised multi-scale attribute-specific localization[C]//Proceedings of the IEEE/CVFInternationalConference onComputerVision. 2019: 4997-5006.

[25]25，Wang K, Dong S, Liu N, et al. PA-Net: Learning localfeatures using by pose attention for short-termperson re-identification[J].Information Sciences, 2021, 565: 196-209.

[26]Huang G, Lu Z, Pun C M, et al. Flexible gait recognition based on flow regulation of local features between key frames[J].IEEEAccess, 2020, 8: 75381-75392.

[27]ChenW C,Yu XY, Ou LL. Pedestrian attribute recognition in video surveillance scenarios based on view-attribute attention localization[J]. Machine Intelligence Research, 2022, 19(2): 153-168.

[28]Chen Z, Li A, Wang Y. A temporal attentive approach for video-based pedestrian attribut recognition[C]//Pattern Recognition and Computer Vision: Second Chinese Conference, PRCV 2019, Xi’an, China, November 8–11, 2019, Proceedings, Part II 2. Springer International Publishing, 2019: 209-220. [29]Hayakawa J, Dariush B. Recognition and 3d localization of pedestrian actions from monocular video[C]//2020 IEEE 23rd International Conference on Intelligent Transportation Systems(ITSC).IEEE, 2020: 1-7.

[30]Hu X,Wei D,Wang Z, et al.Hypergraph video pedestrian re-identification based on posture structure relationship and action constraints[J]. Pattern recognition, 2021, 111: 107688.

[31]Chai T, Chen Z, Li A, et al. Video person re-identification using attribute-enhanced features[J].IEEETransactions onCircuits and Systems for Video Technology, 2022, 32(11):7951-7966.

[32]Sri Preethaa K R, SabariA. Intelligent video analysis for enhanced pedestrian detection byhybridmetaheuristic approach[J]. SoftComputing, 2020, 24(16): 12303-12311.

[33]Wong P K Y, Luo H, Wang M, et al. Recognition of pedestrian trajectories and attributes with computer vision and deep learning techniques[J]. Advanced Engineering Informatics, 2021, 49: 101356.

[34]Mou Q, Wei L, Wang C, et al. Unsupervised domain-adaptive scene-specific pedestrian detection forstatic video surveillance[J]. PatternRecognition, 2021, 118: 108038..

[35]Li Y, Shi F, Hou S, et al. Feature pyramid attention model and multi-label focal loss for pedestrian attribute recognition[J].IEEEAccess, 2020, 8: 164570-164579.

[36]侯瑞兵,常虹,马丙鹏,等.基于时序多尺度互补特征的视频行人重识别[J].计算机学

报,2023,46(01):31-50.

[37]云霄,宋凯莉,张晓光,等.基于时空轨迹融合的遮挡视频行人重识别[J/OL].激光与光电子学进展:1-16[2023-03-22].

[38]姚英茂,姜晓燕.基于图卷积与自注意力图池化的视频行人重识别方法[J/OL].计算机

应用:1-11[2023-03-22].

[39]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/OL]. 煤炭科学技术:1-12[2023-03-22].

[40]葛雯, 姜添元. 改进 YOLO 与 Deepsort 检测跟踪算法的研究[J].计算机仿真,2022,39(05):186-190.

[41]黄凯文,凌六一,王成军,等.基于改进YOLO 和DeepSORT 的实时多目标跟踪算法[J]. 电子测量技术,2022,45(06):7-13.

[42]张伟,沈明霞,刘龙申,等.基于CenterNet 搭配优化DeepSORT 算法的断奶仔猪目标跟

踪方法研究[J].南京农业大学学报,2021,44(05):973-981

[43]Zheng Z, Yang X, Yu Z, Zheng L, YangY and Kautz J.Joint discriminative and generative learning for person reidentification[C]. Computer Vision and Pattern Recognition, 2019, 2138–2147.

[44]ZOPH B,CUBUK E D,GHIASI G,et al.Learning data augtmentation strategies for object det- ection[J].arXiv:1906.11172,2019.

[45]Redmon J,Divvala S,Girshick R,et al.You only look once:unified,realtime object detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR),June27 -30,2016,LasVegas,NV,USA.NewYork:IEEEPress, 2016.779-788.

[46]REDMON J,FARHADI A.YOLO9000:better faster, stronger[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition(CVPR),July 21-26,2017,HonHonolulu, HI,USA. NewYork:IEEEP-ress,2017:6517-6525.

[47]REDMON J,FARHADIA.YOLOv3;an incremental improvement [EB/OL]. (2018-04-08)

[2020-10-05].

[48]GirshickR, Donahue J, DarrellT, et al.Rich feature hierarchiesfor accurate object detection and semantic segmentation. Proceedings of the 2014 IEEE Conference onComputer Vision and PatternRecognition.Columbus:IEEE,2014. 580–587.

[49]Girshick R. Fast R-CNN. Proceedings of the 2015 IEEE International Conference on ComputerVision. Santiago:IEEE, 2015. 1440–1448.

[50]He KM, Gkioxari G, Dollár P, et al. Mask R-CNN.Proceedings of the IEEE International Conference onComputerVision.Venice:IEEE, 2017. 2980–2988. [51]BEWLEYA, GE Z Y, OTT L, et al. Simple on line and realtime tracking[C ]//2016 IEEE InternationalConference on Image Processing (ICI P).Phoenix: IEEE,2016:3464-3468.

[52]CHOLLET F. Xception: Deep learning with depthwise separable convolutions [C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition(CVPR).2017:1800- 1807.

[53]HU J,SHEN L,SUN G.Squeeze and excitation networks[C]//Proceedings of the IEEE Conference onComputerVision and PatternRecognition.2018: 7132-7141