Development of an AI Dataset for Object Detection at Construction Sites

Description

This paper introduces a dataset for object detection training, which includes 12 types of construction machinery commonly operated at civil engineering sites. To collect this dataset, a housing development construction site in South Korea was selected, and a video collection system was operated over a six-month period. Frames were extracted from the collected video footage on a daily basis, resulting in a total of 87,766 images in the full training dataset. Using the COCO Annotator tool, labels and bounding box annotations were processed, generating a total of 856,485 objects.

Steps to reproduce

To ensure transparency and reproducibility, the following outlines how the data was gathered for this research: Site Selection: A civil engineering construction site in South Korea was chosen for data collection. This site was selected based on ongoing earthmoving activities and the presence of various construction equipment types. Video Collection System Setup: A custom video collection system was designed and installed at the site. High-definition (1920x1080 resolution) cameras were positioned at key vantage points around the site to capture clear, comprehensive footage of equipment in operation. The system operated continuously over a six-month period to capture daily site activities under various weather and lighting conditions. Frame Extraction: Frames were extracted from the collected video footage at regular intervals, resulting in a dataset of images suitable for object detection tasks. This step ensured a diverse dataset covering a wide range of equipment movements, angles, and site conditions. Annotation Process: The extracted frames were processed using the COCO Annotator tool, where bounding boxes were manually drawn around 12 specific types of construction equipment (e.g., excavators, bulldozers, cranes). This step required the assistance of trained annotators to ensure accuracy and consistency across the dataset. Data Format: The annotations were then converted to the COCO (Common Objects in Context) format, which is a widely used standard for AI training datasets. This format includes object labels, bounding box coordinates, and other metadata necessary for training object detection models. Tools and Software: Key software and workflows used in this research include: Video collection system: Custom setup using high-definition cameras. Frame extraction: Automated using a custom script to sample images at regular intervals. COCO Annotator: For manual annotation and bounding box creation. Python-based scripts: For data processing and format conversion to the COCO standard. By following these steps, other researchers can reproduce the dataset creation process for similar object detection tasks in construction or civil engineering environments.

Institutions

Categories

Funding

Licence