摘要
煤矿工人和载人车辆(煤矿人车)的智能识别是视频监控系统的重要组成部分,也是煤矿智能化发展的关键任务。然而,煤矿人车检测场景较为复杂,大型人车检测模型部署在有限的计算设备上难以实现,如何在模型检测性能和检测效率之间取得平衡存在诸多挑战。以贵州地区煤矿视频监控数据集为例,提出了一种基于深度学习和模型压缩技术的轻量级煤矿人车检测模型,该模型精准实时的完成了煤矿人车检测任务,对网络进行瘦身的同时几乎没有损失检测性能。具体来说,在网络模型设计阶段,以YOLOv8s为基线提出了一种名为FCW-YOLO的煤矿人车轻量级检测模型,首先将Faster-Block和坐标注意力和开发到网络的特征提取模块中,设计了一种新颖的C2f-Faster-CA轻量级架构,旨在减少网络的冗余通道同时自适应捕捉全局关键信息;其次,采用了WIOU边界回归损失函数以增加模型对普通质量样本的关注,降低了训练样本不平衡带来的回归误差等问题。在模型压缩阶段,联动剪枝算法对提出的FCW-YOLO模型进行通道级别的稀疏,模型可自动识别不重要的通道并对其进行删减,实现了煤矿人车检测模型二次轻量化设计FCWP-YOLO。在自建的煤矿人车检测数据集上的结果表明,提出的模型参数量,计算量和模型大小分别为2.3 M,4.0 GFLOPs,6.0 MB,对比基线模型分别实现了4.9倍、4.7倍、4.4倍的压缩效果,平均检测精度为88.7%,提高了1.1%,每张图像的处理时间仅为5.6 ms。对比多种轻量级架构和先进的检测模型,该方法精度表现优异,计算成本更低,实时性能更好,为资源受限的煤矿场景提供了一种可行的煤矿人车检测方法,满足煤矿视频监控部署要求,可为煤矿人车智能巡检任务提供实时预警。
Intelligent recognition of coal mine workers and manned vehicles(coal mine pedestrian-vehicles)is an important component of video surveillance systems and a key task in the development of coal mine intelligence.However,the detection scene of coal mine pedestrian-vehicles is complex,and deploying large pedestrian-vehicle detection models on limited computing devices is challenging.Balancing between model detection performance and efficiency poses many challenges.This paper proposes a lightweight coal mine pedestrian detection model based on deep learning and model compression techniques.Taking the coal mine video surveillance dataset in Guizhou region as an example.The model accurately and in real-time completes the task of detecting coal mine pedestrian-vehicles,achieving a balance between model detection performance and efficiency.Specifically,in the network model design phase,a lightweight detection model named FCW-YOLO is proposed based on YOLOv8s as the baseline.Faster-Block and coordinate attention are integrated into the feature extraction module of the network,designing a novel C2f-Faster-CA lightweight architecture to reduce redundant channels of the network while adaptively capturing global key information.Furthermore,the WIOU boundary regression loss function is employed to increase the model's focus on common quality samples,addressing issues such as regression errors caused by imbalanced training samples.In the model compression phase,the proposed FCW-YOLO model undergoes channel-level sparsity through a collaborative pruning algorithm,automatically identifying unimportant channels and reducing them,resulting in the FCWP-YOLO model,achieving secondary lightweight design of the coal mine pedestrian-vehicle detection model.Results on a self-built coal mine pedestrian-vehicle detection dataset show that the proposed model has parameters,computational load,and model size of 2.3 M,4.0 GFLOPs,and 6.0 MB,respectively,achieving compression ratios of 4.9 times,4.7 times,and 4.4 times compared to the baseline model.The average detection accuracy is 88.7%,an improvement of 1.1%,with a processing speed of only 5.6ms per image.Compared to various lightweight architectures and advanced detection models,this method demonstrates excellent accuracy,lower computational costs,and better real-time performance,providing a feasible coal mine pedestrian-vehicle detection method for resourceconstrained coal mine scenarios,meeting the deployment requirements of coal mine video surveillance and enabling realtime alerts for intelligent inspection of coal mine pedestrian-vehicles.
作者
解北京
李恒
栾铮
雷振
李晓旭
李卓
XIE Beijing;LI Heng;LUAN Zheng;LEI Zhen;LI Xiaoxu;LI Zhuo(School of Emergency Management and Safety Engineering,China University of Mining and Technology-Beijing,Beijing 100083,China;Guizhou Institute of Technology,College of Mining Engineering,Guiyang 550003,China;Faculty of Architecture,Civil and Transportation Engineering,Beijing University of Technology,Beijing 100124,China)
出处
《煤炭学报》
北大核心
2025年第2期1393-1408,共16页
Journal of China Coal Society
基金
国家重点研发计划资助项目(2022YFC2904100,2023YFC3011300)
中央高校基本科研业务费专项资金资助项目(2023ZKPYAQ04)。
关键词
煤矿工人检测
煤矿载人车辆检测
深度学习
模型压缩
轻量级架构
coal miners detection
coal manned vehicles detection
deep learning
model compression
lightweight architecture
作者简介
解北京(1984-),男,安徽滁州人,副教授,博士。E-mail:bjxie1984@163.com;通讯作者:李恒(1997-),男,贵州盘州人,博士研究生。E-mail:18084282133@163.com。
