Granular stockpile volume dataset.

Area of Science:

• Geospatial Science
• Computer Vision
• Robotics

Background:

• Unmanned Aerial Systems (UAS) are increasingly used for vision-based volume measurements, improving accuracy and automation.
• A lack of comprehensive datasets hinders research on how data collection parameters affect UAS-based volume measurement outcomes.
• Existing research needs standardized data for developing and validating algorithms for stockpile analysis.

Purpose of the Study:

• To introduce a novel, annotated Unmanned Aerial Systems (UAS) dataset for vision-based volume measurement of granular stockpiles.
• To provide researchers with data to investigate the impact of various data collection parameters on 3D model quality and measurement accuracy.
• To facilitate the development of autonomous 3D deep learning models for object detection and measurement.

Main Methods:

• Collected 1521 images of 47 irregular stockpiles (sand, gravel) using UAS in Grand Forks, ND.
• Varied data collection parameters including weather conditions, stockpile size, camera angles, flight patterns, heights, and image overlaps.
• Generated 3D models using Pix4D photogrammetry, annotated point clouds (PLY, XYZ), and linked them with 2D images.

Main Results:

• Generated 3D models with stockpile volumes ranging from 51 m³ to 3000 m³.
• Created an annotated dataset including stockpiles and irrelevant objects (trees, vehicles, roads).
• The dataset comprises unique 3D point data with corresponding 2D images, suitable for deep learning applications.

Conclusions:

• The introduced annotated UAS dataset is a valuable resource for advancing vision-based stockpile volume measurement.
• This dataset supports research into optimizing UAS data collection strategies for enhanced accuracy and automation.
• The dataset is well-suited for training 3D deep learning models for autonomous object detection and measurement in geospatial applications.