基于目标检测的施工钢材物料智能实时计数

Intelligent Real-Time Counting of Construction Materials Based on Object Detection

摘要:

研究了基于深度学习目标检测技术的钢材智能计数方法。通过拍摄并标注大量施工现场的钢筋、圆钢管、方钢管图片，构建了包含近40万个计数点的数据集。基于YOLOv4目标检测算法，建立了钢材智能计数模型，并通过改进其网络结构、损失函数，以及采用合适的训练策略，提高了模型对于钢材的计数精度。检验表明，模型的平均精度为91.41%，平均绝对误差为4.07。利用上述成果开发的APP软件，可通过手机拍照、上传，完成实时计数。

Abstract:

The intelligent counting method of steel based on the deep learning target detection technology is studied in this paper. First， many pictures of steel bars， round steel pipe， and square steel pipe at the construction site are obtained and marked， and a data set containing nearly 400 000 counting points is constructed. Based on the YOLOv4 object detection algorithm， an intelligent steel counting model is established， and the counting accuracy of the steel model is improved by improving its network structure， loss function， and adopting appropriate training strategies. The maximum average precision of the model is 91.41%， with a mean absolute error of 4.07. Based on the above achievements， an APP software is developed， and the real-time counting is completed by taking photos and uploading them.