Integrated 3D Structural Element and Damage Identification: Dataset and Benchmarking

Traditional bridge inspection is a manually performed visual process that is time-consuming, costly, and requires significant support from equipment and resources. Recent advances in artificial intelligence (AI) have accelerated the advances toward inspection automation enabled by robotic imaging and machine vision. Prolific studies in recent literature have reported that damage types and their locations can be identified in 2-dimensional (2D) images with complex scenes using deep learning (DL) based techniques (e.g., semantic object detection or segmentation). However, these efforts have not achieved a level of applicability as meaningful as those from traditional human-based bridge inspection. To enable practical applicability, structural elements and damage patterns must be identified (including detection, localization, and quantification) in a 3D space. To this end, one significant research challenge is the lack of 3D databases for learning both structural elements and damage patterns. This study first developed a unique 3D dataset based on a real bridge structure, and a low-cost LiDAR-enabled imaging device (Intel RealSense) was adopted during the data collection process. Semantic annotations were added for structural elements and damage based on the point clouds and the associated RGB imagery data. Furthermore, a DL-based method was developed to benchmark the usefulness and validity of the dataset. The proposed dataset and DL methods will be open-sourced and expected to facilitate the advances toward engineering inspection automation for bridge structures.