This project aims to create a framework of training engineers and policy makers on robotic operation and image analysis for the inspection and preservation of transportation infrastructure. Specifically, it develops the method for collecting camera-based bridge inspection data and the algorithms for data processing and pattern recognitions; and it creates tools for assisting users on visually analyzing the processed image data and recognized patterns for inspection and preservation decision-making.

The project first developed a Siamese Neural Network to support bridge engineers in analyzing big video data. The network was initially trained by one-shot learning and is fine-tuned iteratively with human in the loop. Bridge engineers define the region of interest initially, then the algorithm retrieves all related regions in the video, which facilitates the engineers to inspect the bridge rather than exhaustively check every frame of the video. Our neural network was evaluated on three bridge inspection videos with promising performances.

Then, the project developed an assistive intelligence system to facilitate inspectors efficiently and accurately detect and segment multiclass bridge elements from inspection videos. A Mask Region-based Convolutional Neural Network was transferred in the studied problem with a small initial training dataset labeled by the inspector. Then, the temporal coherence analysis was used to recover false negative detections of the transferred network. Finally, self-training with a guidance from experienced inspectors was used to iteratively refine the network. Results from a case study have demonstrated that the proposed method uses just a small amount of time and guidance from experienced inspectors to successfully build the assistive intelligence system with an excellent performance.