水运工程施工设备目标检测方法研究  

Research on Object Detection Method for Construction Equipment of Waterway Engineering

作　　者：石兴勇 李志明 冯冬颖 汪承志[2] Shi Xingyong;Li Zhiming;Feng Dongying;Wang Chengzhi(Guangxi Nahai Communications Design&Consultancy Co.,Ltd.,Nanning Guangxi 100084,China;College of Harbor,Waterway and Coastal Engineering,Chongqing Jiaotong University,Chongqing 400074,China;Bahrain Right Banner Water Conservancy Bureau of Chifeng City,Neimengu Bahrain Right Banner 025150,China)

机构地区：[1]广西纳海交通设计咨询有限公司,广西南宁100084 [2]重庆交通大学河海学院,重庆400074 [3]赤峰市巴林右旗水利局,内蒙古赤峰市巴林右旗025150

出　　处：《港口航道与近海工程》2023年第4期76-81,共6页Port,Waterway and Offshore Engineering

摘　　要：本文在YOLOv3(You Only Look Once)基础上提出一种优化的施工机械多目标检测模型,通过引入包含浅层语义的4个特征层的双向特征金字塔方法优化检测模型特征融合结构,提高施工现场多尺度目标检测精确率,并使用考虑边界框中心点距离指标的DIoU(Distance IoU)损失函数方法进一步提升回归框定位效果。结果表明,所提出模型在自建的复杂施工场景数据集上取得了优于原YOLOv3目标检测模型的结果,在检测速度达到实时的0.020 s/帧的情况下,将检测的平均检测准确率(mean Average Percision,mAP)提高了4.5%。本研究方法适用于实际复杂施工场景下的工人及机械多目标检测。Based on YOLOv3(You Only Look Once),an improved multi-object detection model for construction equipment is proposed.By introducing bi-directional feature pyramid network(BiFPN)fusion module based on four feature levels containing low-level features,the detection accuracy of multi-scale objects is improved in construction.Finally,object bounding boxes get better regression by using DIoU loss function.Experimental results show that the proposed model achieves better detection performance than original YOLOv3 model in gathering data on self-built construction scenes,improves mean Average Precision(mAP)by 4.5%based on a real-time detection speed of 0.020 s/frame.The above method is suitable for multi-object detection of workers and equipment under some complex construction conditions.

关 键 词：目标检测 施工监控图像 YOLOv3模型 人工智能 

分 类 号：TU714[建筑科学—建筑技术科学] TU713