Computer Vision Datasets for Visual Blockage Assessment at Culverts

Blockage of culverts caused by transported debris is a major factor in causing flash floods in urban areas. Traditional hydraulic models have been unsuccessful in solving this problem due to a lack of data on peak flood hydraulics and the complex behavior of debris at culverts. To address this problem, a new approach of developing intelligent video analytics (IVA) algorithms is being proposed, which uses computer vision algorithms to extract information about visual blockage. This approach is expected to help in timely and safe maintenance operations and reduce the risk of culverts being blocked. To support the development of computer vision solutions, two datasets have been created: the Synthetic Images of Culvert (SIC) and the Visual Hydraulics Lab Dataset (VHD).

The Synthetic Images of Culvert (SIC) dataset consists of synthetic images of culverts that were generated using a 3D computer application built on the Unity3D gaming engine. The application was designed to simulate various blockage scenarios by allowing users to place different types of debris materials in the scene in various orientations and locations. These blockage scenarios were captured as images using batch capture functionality. The dataset offers diversity in terms of the type of debris (urban, vegetative, mixed), culvert types (pipe, single circular, double circular, single box, double box, triple box), camera viewpoints, time of day, and water levels. However, it has some limitations, such as a single natural background and unrealistic effects and animations.

The Visual Hydraulics-Lab Dataset (VHD) is a dataset of simulated images of culverts that were captured during controlled hydrology lab experiments. The experiments involved a series of tests using scaled physical models of culverts under different flooding conditions. The experiments were recorded using two high definition (HD) cameras and images of culverts in both blocked and unblocked conditions were extracted. The VHD dataset includes a variety of images with different culvert configurations (single circular, double circular, single box, double box), blockage types (urban, vegetative, mixed), simulated lighting conditions, camera viewpoints, and flood levels controlled by inlet water discharge. The limitations of the dataset include reflections from the water surface and flume walls, an identical background and scaling, and clear water.