基于CenterNet−GhostNet的选煤厂危险区域人员检测

张翼翔; 林松; 李雪

doi:10.13272/j.issn.1671-251x.2021080058

基于CenterNet−GhostNet的选煤厂危险区域人员检测

doi: 10.13272/j.issn.1671-251x.2021080058

中国矿业大学(北京) 机电与信息工程学院, 北京　100083

基金项目: 国家自然科学基金资助项目（61936008）。

详细信息

作者简介:
张翼翔（1996—），男，山东泰安人，硕士研究生，研究方向为深度学习，E-mail:2627421136@qq.com

中图分类号: TD67
计量
- 文章访问数: 221
- HTML全文浏览量: 83
- PDF下载量: 53
- 被引次数: 0
出版历程
- 收稿日期: 2021-08-21
- 修回日期: 2022-04-02
- 网络出版日期: 2022-03-05

Personnel detection in dangerous area of coal preparation plant based on CenterNet-GhostNet

School of Mechanical Electronic and Information Engineering, China University of Mining and Technology(Beijing), Beijing 100083, China

摘要

摘要: 选煤厂危险区域人员全身目标检测因粉尘、雾气干扰难以准确与生产环境背景区分，而人员头部特征相对易于辨识，且人头在监控视角下被遮挡的可能性较低，因此危险区域人员检测使用人员头部检测代替人员全身目标检测。目前基于深度学习的轻量化目标检测模型在特征提取时信息损失多，对人头目标的检测能力有限。针对该问题，提出了轻量化人员检测模型CenterNet−GhostNet。该模型以CenterNet网络为基础框架，将轻量化网络GhostNet与特征金字塔相结合作为特征提取网络，GhostNet对输入图像进行特征提取，提升网络特征表达能力，特征金字塔对GhostNet提取的不同分辨率的特征图中包含的信息进行融合，在提取高层语义特征的同时保留较多的细节信息；在较高分辨率的单个输出特征图上使用3个相互独立的卷积操作分支进行解码计算，以充分利用特征图包含的细节信息。实验结果表明：CenterNet−GhostNet模型对佩戴安全帽和未佩戴安全帽两类人头目标的检测精度分别为93.7%和91.7%，均优于通用的轻量化模型SSD−MobileNet、YOLOv4 Tiny和CenterNet−Res18；CenterNet−GhostNet模型部署在NVIDIA Jetson Nano上的单帧检测耗时为67 ms，满足选煤厂危险区域人员高精度、实时检测要求。
- 选煤厂 /
- 危险区域 /
- 人员检测 /
- 目标检测 /
- 轻量化网络 /
- 深度学习
Abstract: Due to the dust and fog interference, it is difficult to distinguish accurately the whole body target of personnel in dangerous areas of coal preparation plant from the production environment background. Moreover, the head features of personnel are relatively easy to be identified, and the possibility of head being blocked in the monitoring perspective is low. Therefore, the head detection of personnel in dangerous areas is used instead of the whole body target detection of personnel. At present, the lightweight target detection model based on deep learning has a lot of information loss in feature extraction, and its detection capability of human head target is limited. In order to solve this problem, a lightweight personnel detection model CenterNet-GhostNet is proposed. The model takes CenterNet network as the basic framework, and combines the lightweight network GhostNet and the feature pyramid network(FPN) as the feature extraction network. GhostNet extracts the features of the input image and improves the network feature expression capability. And the FPN fuses the information contained in the feature maps with different resolutions extracted by GhostNet, so that more detailed information is reserved while extracting the high-level semantic features. Three independent convolution operation branches are used to decode and calculate the single output feature map with higher resolution, so as to make full use of the detailed information contained in the feature map. The experimental results show that the detection precision of CenterNet-GhostNet model is 93.7% and 91.7% respectively for the two types of head targets with and without helmet, which are better than the general lightweight models SSD-MobileNet, YOLOv4 Tiny and CenterNet-Res18. The single frame detection time of CenterNet-GhostNet model deployed on NVIDIA Jetson Nano is 67 ms, which meets the requirements of high-precision and real-time detection of personnel in dangerous areas of coal preparation plant.
- coal preparation plant /
- dangerous area /
- personnel detection /
- target detection /
- lightweight network /
- deep learning

HTML全文

图 1 CenterNet−GhostNet模型框架

Figure 1. Framework of CenterNet-GhostNet model

下载: 全尺寸图片幻灯片

图 2 基于GhostNet和FPN的特征提取网络结构

Figure 2. Feature extraction network structure based on GhostNet and feature pyramid network

下载: 全尺寸图片幻灯片

图 3 G−bneck结构

Figure 3. Structure of G-bneck

下载: 全尺寸图片幻灯片

图 4 Ghost模块

Figure 4. Ghost module

下载: 全尺寸图片幻灯片

图 5 CenterNet−GhostNet模型检测结果

Figure 5. Detection results of CenterNet-GhostNet model

下载: 全尺寸图片幻灯片

图 6 CenterNet−Res18模型检测结果

Figure 6. Detection results of CenterNet-Res18 model

下载: 全尺寸图片幻灯片

图 7 选煤厂危险区域人员检测布置

Figure 7. Personnel detection arrangement in dangerous area of coal preparation plant

下载: 全尺寸图片幻灯片

表 1 GhostNet参数

Table 1. Parameters of GhostNet

输入尺寸	结构	步长
512×512×3	Conv2d（3×3）	2
256×256×16	G−bneck×2	1，2
128×128×24	G−bneck×2	1，2
64×64×40	G−bneck×2	1，2
32×32×80	G−bneck×4	1，1，1，1
32×32×112	G−bneck×2	1，2
16×16×160	G−bneck×4	1，1，1，1
16×16×160	Conv2d（3×3）	1

下载: 导出CSV

表 2 不同模型检测精度和检测速度对比

Table 2. Comparison of detection accuracy and detection speed among different models

模型	mAP/%		FPS/（帧·s⁻¹）
模型	佩戴安全帽	未佩戴安全帽	FPS/（帧·s⁻¹）
SSD−MobileNet	87.1	83.3	167
YOLOv4 Tiny	91.5	89.7	117
CenterNet−Res18	88.6	85.1	155
CenterNet−GhostNet	93.7	91.7	146

下载: 导出CSV

参考文献(14)

[1]	张宗华. 选煤厂人员智能视频监控系统设计[J]. 工矿自动化,2013,39(4):76-79. doi: 10.7526/j.issn.1671-251X.2013.04.020 ZHANG Zonghua. Design of intelligent video monitoring system of personnel of coal preparation plant[J]. Industry and Mine Automation,2013,39(4):76-79. doi: 10.7526/j.issn.1671-251X.2013.04.020
[2]	张圣强,王建军,丁蕾. 选煤厂智能视频监控中的运动人体目标检测方法[J]. 工矿自动化,2011,37(11):60-62. ZHANG Shengqiang,WANG Jianjun,DING Lei. Detecting method of moving human in intelligent video monitoring of coal preparation plant[J]. Industry and Mine Automation,2011,37(11):60-62.
[3]	陈海龙. 煤矿选煤厂巡检机器人的研究与设计[D]. 徐州: 中国矿业大学, 2014. CHEN Hailong. The study and design of coal cleaning plant inspection robot[D]. Xuzhou: China University of Mining and Technology, 2014.
[4]	林娅静. 基于视觉的选煤厂智能监控系统研究[D]. 徐州: 中国矿业大学, 2014. LIN Yajing. Research on intelligent monitoring system of coal preparation plant based on vision[D]. Xuzhou: China University of Mining and Technology, 2014.
[5]	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]. 徐州: 中国矿业大学, 2019. ZHOU Chenhui. Research on personnel detection and statistical analysis in coal mine complex scenes based on deep learning[D]. Xuzhou: China University of Mining and Technology, 2019.
[6]	ZHOU Xingyi, WANG Dequan, KRAHENBUHL P. Objects as points[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, 2019: 7263-7271.
[7]	HAN K, WANG Yunhe, TIAN Qi, et al. GhostNet: more features from cheap operations[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Seattle, 2016: 1577-1586.
[8]	LIN T Y, DOLLAR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, 2017: 2117-2125.
[9]	HOWARD A G, ZHU Menglong, CHEN Bo, et al. MobileNets: efficient convolutional neural networks for mobile vision applications[EB/OL]. [2021-08-02]. https://arxiv.org/abs/1704.04861.
[10]	DAI Jifeng, QI Haozhi, XIONG Yuwen, et al. Deformable convolutional networks[C]//IEEE International Conference on Computer Vision, Venice, 2017: 764-773.
[11]	XU Yuanyuan,YAN Wan,YANG Genke,et al. CenterFace:joint face detection and alignment using face as point[J]. Scientific Programming,2020,2020:1-8.
[12]	LIN T Y,GOYAL P,GIRSHICK R,et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2020,42(2):318-327. doi: 10.1109/TPAMI.2018.2858826
[13]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: optimal speed and accuracy of object detection[EB/OL]. [2021-08-02]. https://arxiv.org/abs/2004.10934.
[14]	EVERINGHAM M,GOOL L V,WILLIAMS C K I,et al. The PASCAL visual object classes( VOC) challenge[J]. International Journal of Computer Vision,2010,88(2):303-338. doi: 10.1007/s11263-009-0275-4