基于深度学习的动态危险区域入侵检测方法

Intrusion detection method of dynamic hazardous area based on deep learning

为解决天车吊装动态危险区域难识别、位置信息难确定、车间生产环境复杂导致特殊的相机视角所获取的人员目标特征信息不明显,影响漏检误检和标准目标检测模型部署应用难等问题,提出1种基于深度学习的动态危险区域人员入侵检测方法。该方法首先利用双目视觉技术完成重物提升判断及天车吊装动态危险区域检测2项任务,然后利用提出的PDConv轻量化卷积模块和卷积注意力机制模块对YOLOv5基础模型进行改进,并设计开发天车吊装动态危险区域人员入侵检测系统。研究结果表明:利用双目视觉技术可以更精准获取动态危险区域位置信息;改进后的YOLOv5基础模型在保证模型检测精度的同时可降低适当参数量;天车吊装动态危险区域人员入侵检测系统可实现远程管控天车危险区域入侵检测任务,并利用振动报警装置可实时对入侵人员进行提醒管理,该系统在实际测试中的漏检误检率为2%,满足工业生产需求。研究结果可为动态危险区域入侵检测研究提供参考。

In order to solve the problems that the dynamic hazardous area of crane hoisting is difficult to recognize,the location information is difficult to determine,the workshop production environment is complex,and the personnel target feature information obtained from the special camera perspective is not obvious,which leads to the problems of missed detection and false detection,and the standard target detection model is difficult to deploy and apply,a detection method for the personnel intrusion of dynamic hazardous area based on deep learning was proposed.The binocular vision technology was applied to complete two tasks of heavy lifting judgment and dynamic hazardous area detection of crane hoisting,then the proposed PDConv lightweight convolution module and convolutional attention mechanism module were utilized to improve the YOLOv5 base model,and a detection system for the personnel intrusion of dynamic hazardous area in crane hoisting was designed and developed.The results show that the binocular vision technology can accurately obtain the location information of dynamic hazardous area,the improved YOLOv5 basic model can reduce the number of appropriate parameters while ensuring the detection accuracy of the model.The detection system for the personnel intrusion of dynamic hazardous area in crane hoisting can realize the remote control on the intrusion detection task of crane hazardous area,and use the vibration alarm device to conduct the remind management of intrusion personnel in real time.In the actual test,the missed detection and false detection rate of the system is 2%,which meets the needs of industrial production.The research results can provide reference for the research on intrusion detection of dynamic hazardous areas.