Automated bridge component recognition from point clouds using deep learning

To assess the current conditions of bridges in practice, a manual visual inspection is commonly used for maintenance purposes. Because manual visual inspection is time consuming, expensive, and laborious, computer vision with deep learning techniques has recently been introduced as a promising tool for automatic damage detection and classification. However, damage localization is difficult in that close-up images do not contain the global structural context, while knowing the damage locations and the associated structural components is essential for interpreting the overall structural health. Compared with two-dimensional (2D) image data, the point clouds in the three-dimensional (3D) space with an extra dimension can be useful for bridge component recognition. However, previous research employing geometric features of bridges was only partially successful for relatively simple types of bridges without the background regions (e.g., ground, water, and vegetation). This study presents a methodology for automated bridge component recognition using deep learning. The proposed approach is designed for general bridges that may have curved decks or different pier heights. Furthermore, the proposed method can handle point clouds that have points in the background regions, significantly reducing the time-consuming preprocessing of the point cloud. For robust and automated segmentation, a set of point clouds is extracted from a bridge by subspace partition, and a deep leaning technique is employed to classify labels. Subsequently, the classification results are combined to determine the consensus label for each point based on a majority of estimated classes, thereby improving identification accuracy. The classification performance is experimentally validated using full-scale bridge.