一种基于YOLOv5-Lite算法的船员行为图像识别方法  

Image Recognition of Unsafe Behavior of Seafarers Based on YOLOv5-Lite Algorithm

作　　者：薛正源 韩冰[1] 王炳德 张兴龙[1] XUE Zhengyuan;HAN Bing;WANG Bingde;ZHANG Xinglong(Warship Automatic System Division,Shanghai Ship and Shipping Research Institute Co.,Ltd.,Shanghai 200135,China)

机构地区：[1]上海船舶运输科学研究所有限公司舰船自动化系统事业部,上海200135

出　　处：《上海船舶运输科学研究所学报》2023年第2期33-40,54,共9页Journal of Shanghai Ship and Shipping Research Institute

基　　金：高技术船舶科研项目(工信部重装函[2020]313号:CJ02N20);上海市优秀学术/技术带头人计划项目(22XD1431000)。

摘　　要：为及时发现船员的不安全行为,降低因船员行为不当而引发船舶事故的风险,提出一种基于YOLOv5-Lite算法的船员行为图像识别方法。在YOLOv5算法中引入Fire Module结构,并对其通道进行变换,减少图像识别模型残差模块的数量,减小模型的体量,提升模型的运行效率。以对船上重点作业区域的船员是否佩戴安全帽进行识别为例,建立基于YOLOv5-Lite算法的图像识别模型,收集图像数据,构建HTST(Helmet Tumble Smoke Tired)数据集,对模型进行训练,实现对人员动作的识别。将采用该算法与采用YOLOv3和YOLOv4等算法所得识别结果相比较,验证该算法的有效性。结果表明:基于YOLOv5-Lite算法的船员行为图像识别方法能有效识别船员的不安全行为,识别的准确度能达到95.70%;相比另外2种算法,YOLOv5-Lite算法具有更好的稳定性和检测效果,FPS(Frames Per Second)检测速度快,满足对船员的行为进行实时识别的需求。In order to imediately detect unsafe actions of crew members and to reduce the risk of accidents caused by crew misoperation,a method for identifying unsafe actions of crew members based on YOLOv5-Lite algorithm is proposed.The Fire Module structure is introduced into the YOLOv5 algorithm to achieve channel transformation,which also reduces the number of image recognition model residual modules and the size of the model,and,in turn,to improve the operational efficiency of the model.As an example,an image recognition model for detecting persons in operation area who is not wearing a safety helmet is established.The model collects image data,generates an HTST(Helmet Tumble Smoke Tired) dataset,and performs image processing.The model is trained to achieve the identification of movements of people.The output of the model is compared with the recognition results obtained with algorithms YOLOv3 and YOLOv4 to verify the effectiveness of the method.The results show that the method for identifying unsafe behaviors of crew members based on YOLOv5-Lite algorithm can effectively identify unsafe behaviors of crew members with a mAP value of 95.70%.YOLOv5-Lite algorithm has better stability and detection capability than that of YOLOv3 and YOLOv4.The FPS(Frames Per Second) detection speed is fast,meeting the demand for real-time identification of unsafe behavior of crew members.

关 键 词：安全行为识别 YOLOv5-Lite算法 视频监控 图像识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]