Dataset and benchmark for detecting moving objects in construction sites

• Published in: Automation in construction Vol. 122; p. 103482
• Main Authors: Xuehui, An, Li, Zhou, Zuguang, Liu, Chengzhi, Wang, Pengfei, Li, Zhiwei, Li
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.02.2021; Elsevier BV
• Subjects: Artificial neural networks; Benchmark; Benchmarks; Construction; Construction site; Construction sites; Control equipment; Dataset; Datasets; Deep neural networks; Detectors; Moving object recognition; Object detection; Quality control; Statistical analysis; Stormwater; Benchmark; Deep neural networks; Construction site; Dataset; Object detection
• ISSN: 0926-5805; 1872-7891
• DOI: 10.1016/j.autcon.2020.103482

Detecting workers and equipment through images/videos can assist in safety monitoring, quality control, and productivity management at construction sites. Currently, the dominant method for detecting is Deep Neural Networks (DNNs). To apply this method, the DNNs always need to be trained on image datasets that contain objects at the construction site. However, a large-scale and publicly available image dataset for detecting objects at construction sites is still absent, and this hinders research in this field. In this study, the Moving Objects in Construction Sites (MOCS) image dataset is presented. The dataset contains 41,668 images collected from 174 different construction sites. Thirteen categories of moving objects found in construction sites were annotated. Furthermore, the objects were precisely annotated using per-pixel segmentations to assist in precise object localization. A detailed statistical analysis was performed in this study. Finally, a benchmark containing 15 different DNN-based detectors was made using the MOCS dataset. The results show that all detectors trained on the dataset could detect objects at construction sites precisely and robustly. •A large-scale image dataset for detecting moving objects in construction sites is developed and made publicly available in this paper.•A benchmark has been made on the dataset to guide others selecting detectors.•Limitations and future improvements of the dataset and benchmark are carefully discussed.