基于深度学习与深度估计的施工机械危险区域侵入智能预警方法

Intelligent early warning method for construction machinery hazardous area intrusion based on deep learning and depth estimation

为解决因工人和施工机械侵入施工危险区域等原因造成的工程安全事故问题,提出一种多任务驱动的施工机械危险区域侵入事件动态识别与预警方法。首先以置换可变性卷积DConv2模块的Yolov8网络进行目标类别检测和坐标外轮廓提取,提高移动施工机械的识别准确率;然后结合Monodepth2单目深度估计网络进行深度信息估计和坐标统一,计算工人或施工机械距离危险区域事件的实际距离,用于评估和预警危险区域侵入风险;最后将测试结果与不同修改层的Yolov8、原始Yolov8和Yolov5模型进行对比,并设计4种场景进行模型性能验证。结果表明:模型在施工机械的识别和轮廓提取精度上分别提高了2.99%和3.55%,对工人和施工机械侵入移动施工机械危险区域风险事件的识别准确率能保持在88%以上,FPS保持在17.7左右,可以有效实现对施工机械危险区域侵入事件的智能动态预警。

To solve the problem of engineering safety accidents caused by workers and construction machinery intruding into construction hazardous areas,etc.,a multi-task-driven dynamic identification and early warning method for hazardous area intrusion events is proposed.A Yolov8 network with permutation variable convolutional DConv2 module was used for target class detection and coordinate outer contour extraction to improve the recognition accuracy of mobile construction machinery.It was also combined with the Monodepth2 monocular depth estimation network for depth information estimation and coordinate unification to calculate the true state of a worker or construction machine at a distance from a hazardous area event,which is used to assess the risk of hazardous area intrusion.The model performance was compared with Yolov8,the original Yolov8 and Yolov5models with different modification layers and four scenarios were designed for model performance validation.The study shows that the model improves 2.99%and 3.55%in construction machinery identification accuracy and contour extraction accuracy respectively,and can maintain an accuracy rate of over 94%in the identification of workers and construction machinery intrusion into mobile construction machinery hazardous area,with an FPS of around 17.7,which can effectively achieve intelligent dynamic warning of construction hazardous area intrusion.