一种面向密集场景的轻量化人群检测网络

doi:10.16180/j.cnki.issn1007-7820.2023.08.006

电子科技 ›› 2023, Vol. 36 ›› Issue (8): 35-42.doi: 10.16180/j.cnki.issn1007-7820.2023.08.006

一种面向密集场景的轻量化人群检测网络

潘昊¹,刘翔¹,赵静文¹,张星^2,³

1.上海工程技术大学电子电气工程学院,上海 201620
2.上海工程技术大学管理学院,上海 201620
3.江苏大学汽车工程学院,江苏镇江 212023

收稿日期:2022-03-19 出版日期:2023-08-15 发布日期:2023-08-14
作者简介:潘昊(1996-),男,硕士研究生。研究方向:行人检测、行人跟踪。|刘翔(1972-),男,博士,副教授。研究方向:医学影像与多智能机器人协同、视频大数据的智能分析与高速检索、缺陷检测与高精度测量等。|赵静文(1992-),女,博士,讲师。研究方向:生物医学图像处理和分析。
基金资助:
中国高校产学研创新基金(2021FNB02001);文化部科技创新项目(2015KJCXXM19)

A Lightweight Crowd Detection Network for Dense Scenes

PAN Hao¹,LIU Xiang¹,ZHAO Jingwen¹,ZHANG Xing^2,³

1. School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China
2. School of Management,Shanghai University of Engineering Science,Shanghai 201620,China
3. School of Automotive and Traffic Engineering, Jiangsu University,Zhenjiang 212023,China

Received:2022-03-19 Online:2023-08-15 Published:2023-08-14
Supported by:
China University Industry-University-Research Innovation Fund(2021FNB02001);Science and Technology Innovation Project of the Ministry of Culture(2015KJCXXM19)

摘要/Abstract

摘要：

针对密集场景下行人检测的遮挡问题,文中提出了基于YOLO(You Only Look Once)的SC-YOLOv4人群检测网络。在YOLOv4的CSPNet(Cross Stage Partial Network)结构基础上,结合ShuffleNetv2网络思想改进普通卷积结构,将原来普通的残差模块替换为Shuffle Module模块,提出了基于S-CSPDarkNet53(Shuffle CSPDarkNet53)的骨干网络结构,在保留精度的同时降低了网络参数量。文中在保留原来PANet(Path Aggregation Network)结构的基础上设计中心点预测模块,将原来的3个输出特征层改用基于中心点的预测方法,即对目标的中心点进行回归和训练计算损失,摒弃了原来的NMS(Non-Maximum Suppression)操作,进一步提高遮挡情况下的检测精度。实验结果表明,在CrowdHuamn数据集上采用S-CSPDarkNet53结构的YOLOv4较原网络的参数量显著减少,检测速度提升了5.2 frame·s^-1,而最终的SC-YOLOv4网络在检测速度上较YOLOv4提升了4.9 frame·s^-1。

关键词: 人群检测, YOLO, Shuffle Module, 中心点检测, 密集人群, CrowdHuman, CSPNet, YOLOv4

Abstract:

For the occlusion problem of pedestrian detection in dense scenes, this study proposes the SC-YOLOv4 crowd detection network based on YOLO. Based on the CSPNet structure of YOLOv4 and combined with the idea of ShuffleNetv2 network, the common convolution structure is improved, and the original common residual module is replaced with the Shuffle Module. A backbone network structure based on S-CSPDarkNet53 is proposed, which preserves the accuracy and reduces the number of network parameters. The centroid prediction module is designed on the basis of retaining the original PANet structure, and the original three output feature layers are replaced with a centroid-based prediction method, that is, the regression and training of the target center point are carried out to calculate the loss, and the original NMS operation is discarded to further improve the detection accuracy in the case of occlusion. The experimental results show that YOLOv4 with S-CSPDarkNet53 structure on the CrowdHuamn data set reduces the amount of parameters and improves the detection speed by 5.2 frame·s^-1 when compared with the original network. Compared with YOLOv4, the final SC-YOLOv4 network improves the detection speed by 4.9 frame·s^-1.

Key words: crowd detection, YOLO, Shuffle Module, center point detection, dense crowd, CrowdHuman, CSPNET, YOLOv4

中图分类号:

TP391.41

潘昊,刘翔,赵静文,张星. 一种面向密集场景的轻量化人群检测网络[J]. 电子科技, 2023, 36(8): 35-42.

PAN Hao,LIU Xiang,ZHAO Jingwen,ZHANG Xing. A Lightweight Crowd Detection Network for Dense Scenes[J]. Electronic Science and Technology, 2023, 36(8): 35-42.

图/表 9

图1

图2

图3

图4

表1

表2

图5

表3

图6

参考文献 21

[1]	Viola P, Jones M J. Rapid object detection using a boosted cascade of simple features[C]. Kauai: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2001:377-390.
[2]	Dollár P, Tu Z, Perona P, et al. Integral channel features[C]. London: British Machine Vision Conference, 2009:7-10.
[3]	Dalal N, Triggs B. Histograms of oriented gradients for human detection[C]. San Diego: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2005:886-893.
[4]	Ren S, He K, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. Advances in Neural Information Processing Systems, 2017, 39(6):1137-1149.
[5]	He K, Gkioxari G, Dollar P, et al. Mask R-CNN[J]. Proceedings of the IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 42(2):386-397.
[6]	Chu X, Zheng A, Zhang X, et al. Detection in crowded scenes: One proposal, multiple predictions[C]. Virtual: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020:12214-12223.
[7]	Zhang S, Wen L, Bian X, et al. Occlusion-aware R-CNN: detecting pedestrians in a crowd[C]. Munich: Proceedings of the Fifteenth European Conference on Computer Vision. Cham: Springer, 2018:657-674.
[8]	Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector[C]. Amsterdam: Proceedings of the IEEE European Conference on Computer Vision, 2016:21-37.
[9]	Redmon J, Farhadi A. Yolov3: An incremental improvement[EB/OL]. (2018-04-08) [2021-9-15] https://arxiv.org/pdf/1804.02767.pdf
[10]	Bochkovskiy A, Wang C, Liao H. Yolov4: Optimal speed and accuracy of object detection[EB/OL]. (2004-10-34) [2021-10-21] https://arxiv.org/pdf/2004.10934.pdf
[11]	Ultralytics. YOLOv5[EB/OL]. (2020-06-09) [2021-9-31] https://github.com/ultralytics/yolov5.
[12]	刘竣文, 张永军, 李智, 等. 基于RDM-YOLOv3的头部检测[J]. 激光与光电子学进展, 2022, 59(8):1-15.
	Liu Junwen, Zhang Yongjun, Li Zhi, et al. Head detection based on RDM-YOLOv3[J]. Laser & Optoelectronics Progress, 2022, 59(8):1-15.
[13]	He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]. Las Vegas: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016:770-778.
[14]	Huang G, Liu Z, Van D M L, et al. Densely connected convolutional networks[C]. Honolulu: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017:4700-4708.
[15]	蔡逢煌, 张岳鑫, 黄捷. 基于YOLOV3与注意力机制的桥梁表面裂痕检测算法[J]. 模式识别与人工智能, 2020, 33(10):926-933. doi: 10.16451/j.cnki.issn1003-6059.202010007
	Cai Fenghuang, Zhang Yuexin, Huang Jie. Bridge surface crack detection algorithm based on YOLOv3 and attention mechanism[J]. Pattern Recognition and Artificial Intelligence, 2020, 33(10):926-933. doi: 10.16451/j.cnki.issn1003-6059.202010007
[16]	叶飞, 刘子龙. 基于改进YOLOv3算法的行人检测研究[J]. 电子科技, 2021, 34(1):5-9.
	Ye Fei, Liu Zilong. Research on pedestrian detection based on improved YOLOv3 algorithm[J]. Electronic Science and Technology, 2021, 34(1):5-9.
[17]	Law H, Deng J. CornerNet: Detecting objects as paired keypoints[J]. International Journal of Computer Vision, 2020, 128(3):642-656. doi: 10.1007/s11263-019-01204-1
[18]	Zhou X, Wang D, Krähenbühl P. Objects as points[C]. Long Beach: IEEE Conference on Computer Vision and Pattern Recognition, 2019:35-43.
[19]	Ma N, Zhang X, Zheng H T, et al. Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]. Munich: Proceedings of the European Conference on Computer Vision, 2018:116-131.
[20]	Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[C]. Venice: Proceedings of the IEEE International Conference on Computer Vision, 2017:2980-2988.
[21]	Ge Z, Liu S, Wang F, et al. Yolox: Exceeding yolo series in 2021[EB/OL]. (2021-08-06) [2021-10-30] https://arxiv.org/pdf/2107.08430.pdf.

人数/图片	Caltech		KITTI		CityPerson		COCOPerson		CwordHuman
1	7 839	18.3%	969	26.1%	2 482	83.4%	64 115	100.0%	15 000	100.0%
2	3 257	7.6%	370	10.0%	2 082	70.0%	39 283	61.3%	15 000	100.0%
3	1 265	3.0%	273	7.4%	1 741	58.5%	28 553	44.5%	14 996	100.0%
5	282	0.7%	164	4.4%	1 225	41.2%	18 775	29.3%	14 220	94.8%
10	36	0.1%	19	0.5%	610	20.5%	9 604	15.0%	10 844	72.3%
20	0	0.0%	0	0.0%	227	7.6%	0	0.0%	5 907	39.4%
30	0	0.0%	0	0.0%	94	3.2%	0	0.0%	3 294	21.9%

网络模型	模型权重 /MB	模型参数量/MB	FPS	AP /%
YOLOv4	244.4	244.3	59.8	42.9
YOLOv4(S-CSPDarkNet53)	161.7	162.2	65.0	41.5

网络模型	AP/%	MR	FPS
SSD(VGG-16)	30.5	0.64	63.5
Faster-RCNN(resnet50)	39.1	0.58	15.3
CenterNet(Hourglass)	43.7	0.54	54.1
YOLOv3	32.8	0.65	54.6
YOLOv4	42.9	0.58	59.8
YOLOv5	44.1	0.53	70.6
YOLOX	46.5	0.47	66.5
YOLOv4(point based)	45.9	0.48	60.0
YOLOv4(S-CSPDarkNet53+point based)	44.3	0.49	64.7

一种面向密集场景的轻量化人群检测网络

A Lightweight Crowd Detection Network for Dense Scenes

RichHTML

PDF (PC)

赞

可视化

摘要/Abstract

引用本文

使用本文

图/表 9

参考文献 21

相关文章 7

Metrics

本文评价

推荐阅读 0

[1]	胡永阳,李淼,孟凡开,张峰,孟艺薇,宋宇鲲. 基于Tiny-YOLOv3的网络结构化压缩与加速[J]. 电子科技, 2023, 36(8): 43-48.
[2]	郑玉恒,付东翔. 基于Slim-YOLOv4与嵌入式设备的无人机检测[J]. 电子科技, 2023, 36(5): 55-61.
[3]	唐政,张会林,马立新,刘金芝,王昊. 轻量级网络算法对输电线路上异物目标的识别[J]. 电子科技, 2023, 36(4): 71-77.
[4]	程长文,陈玮,陈劲宏,尹钟. 改进YOLO的口罩佩戴实时检测方法[J]. 电子科技, 2023, 36(2): 73-80.
[5]	张乔木,钟倩文,孙明,罗文成,柴晓冬. 复杂环境下弓网接触位置动态监测方法研究[J]. 电子科技, 2022, 35(8): 66-72.
[6]	赵轩,周凡,余汉成. 基于改进特征提取及融合模块的YOLOv3模型[J]. 电子科技, 2022, 35(7): 40-45.
[7]	叶飞, 刘子龙. 基于改进YOLOv3算法的行人检测研究[J]. 电子科技, 2021, 34(1): 5-9.