面向施工机械的深度学习图像数据集合成方法

工程现场环境复杂,获取包含丰富信息的图像难度大且标注成本高,造成基于计算机视觉的深度学习施工机械图像数据集构建困难。为满足快速、高质量构建建筑工程领域施工机械深度学习图像数据集,提出一种基于三维建模引擎的施工机械图像生成与自动标注方法,并以挖掘机为例构建了名为SCED(Synthesized Construction Equipment Dataset)的挖掘机数据集。首先,采用三维建模引擎UE4对目标挖掘机设备进行模型构建,然后借助UnrealCV工具对原始模型进行多角度、多区域的图像采集,使用自编写模块实现自动语义分割与掩码图像生成,并完成图像的自动标注,最终生成包含10 000张图像的数据集。与现有公开机械数据集进行了目标尺寸、数量与构建工作量的对比,并比较了构建效率与成本,最后进行了图像数据集质量与效果验证。结果表明:该构建方法综合效率更高且成本更低,构建的SCED图像数据集丰富性和泛化能力更好,针对小目标物具有更好的检测效果; 研究成果可为今后建筑施工领域深度学习图像数据集的构建提供参考依据。

Image Dataset Synthetic Method for Construction Machinery Based on Deep Learning

The complex environment of construction sites, the difficulty of obtaining images containing rich information and the high cost of image annotation make it difficult to build deep learning construction machinery image datasets based on computer vision. In order to build a fast and high-quality deep learning image dataset of construction machinery in the construction field, a method of generating and automatically annotating images of construction machinery based on a 3D modeling engine was proposed, and an excavator dataset named SCED(Synthesized Construction Equipment Dataset)was constructed with an excavator as an illustration. Firstly, the target excavator equipment was modelled using the 3D modelling engine UE4, then the original model was captured from multiple angles and regions utilizing UnrealCV tool, automatic semantic segmentation and mask image generation were achieved using a self-written module, and the bounding boxes of excavator instance were automatically annotated, generating a dataset containing 10 000 images. The target size, number and construction effort were compared with existing publicly available mechanical datasets, and the construction efficiency and cost were also compared, and finally the quality and effectiveness of the image dataset were verified. The results show that the construction method is more efficient and less expensive, and the SCED image dataset is richer and better generalised, with better detection results for small targets. The research findings can provide a reference for the construction of deep learning image datasets in the field of building construction in the future.