【摘要】 施工现场光照多变、背景复杂、施工人员形态多样,给安全帽佩戴情况检测带来很大的困难。针对传统检测方法准确率低、鲁棒性差的问题,本文提出了一种基于深度学习的安全帽佩戴情况检测方法。该方法以YOLOv2目标检测方法为基础,对其网络结构进行了改进。首先借鉴了密集连接网络思想,在原网络中加入了密集块,实现了多层特征的融合以及浅层低语义信息与深层高语义信息的兼顾,提高了网络对于小目标检测的敏感性;然后,利用MobileNet中的轻量化网络结构对网络进行压缩,使模型的大小缩减为原来的十分之一,增加了模型的可用性。采用自制的HelmetWear数据集对改进后的网络模型进行训练和测试,并将该模型与原YOLOv2和最新的YOLOv3进行了对比,结果显示:该模型的检测准确率为87.42%,稍逊色于YOLOv3,但是其检测速度提升显著,比YOLOv2和YOLOv3分别提高了37%和215%,可达148frame/s。实验表明,改进后的网络模型能在保证检测准确率的同时,有效减小参数量,显著提升检测速度。更多还原

【Abstract】 Construction sites comprise changeable illumination,complicated backgrounds,and various types of construction personnel,which makes the detection of helmet wearing challenging.To address the problems of low accuracy and poor robustness of traditional detection methods,this paper proposed a method of helmet wearing detection based on deep learning.The proposed method was based on the YOLOv2 target detection,and its network structure was improved.First,utilizing the notion of densely connected networks,dense blocks are added to the original network,which aids in the realization of the fusion of multi-layer features and the combination of shallow low semantic information and deep high semantic information,thereby improving the network sensitivity to enable the detection of small targets.Subsequently,the lightweight network structure in MobileNet was used for network compression,thus reducing the size of the model to one tenth of its original and also increasing model availability.The improved network model was trained and tested on the self-made HelmetWear dataset and compared with the original as well as the latest YOLOv3.The obtained results show that the detection accuracy of the model is 87.42%,which is slightly lower than that of YOLOv3,but its detection speed is significantly improved,i.e.,37% and 215% higher than that of YOLOv2 and YOLOv3,respectively,while reaching 148 frames/s.Experiments confirm that the proposed model can effectively reduce parameter quantity and significantly enhance detection speed while ensuring detection accuracy.更多还原