Physics-Based Graphics Models in 3D Synthetic Environments as Autonomous Vision-Based Inspection Testbeds.

Area of Science:

• Civil Engineering
• Computer Vision
• Robotics

Background:

• Manual infrastructure inspection is risky, subjective, and slow.
• Autonomous inspection using deep learning and robots is advancing but lacks integrated testing.
• Real-world data for diverse inspection scenarios is challenging to acquire.

Purpose of the Study:

• To propose a framework for a virtual visual inspection testbed using 3D synthetic environments.
• To enable end-to-end testing of autonomous inspection strategies.
• To address the concurrent needs of data acquisition, processing, and decision-making in autonomous inspection.

Main Methods:

• Developed a framework utilizing physics-based graphics models (PBGMs) for realistic damage rendering in 3D synthetic environments.
• Created a large-scale annotated dataset, QuakeCity, for post-earthquake building inspections.
• Conducted experiments using PBGMs to generate data and validate deep learning models.

Main Results:

• Generated the QuakeCity dataset, enabling large-scale annotated benchmark data for post-earthquake inspections.
• Demonstrated improved performance of deep learning models trained on the QuakeCity dataset for real-world data inference.
• Validated PBGMs as an effective testbed for developing and validating autonomous vision-based inspection strategies.

Conclusions:

• The proposed virtual testbed facilitates robust end-to-end testing of autonomous inspection systems.
• Physics-based graphics models (PBGMs) are crucial for generating realistic synthetic data for training and validation.
• This approach enhances the development and reliability of autonomous vision-based inspections for civil infrastructure.