Synchronized identification and localization of defect on the bottom of steel box girders based on a dynamic visual perception system

Abstract

Inspecting the underside of large-span bridges is a major challenge due to the extensive area and inaccessibility. This study developed a system that integrates advanced equipment with intelligent algorithms, designed to achieve precise identification and rapid localization of defects on the underside of bridges. The key components of the system are summarized as follows: (1) The dynamic visual perception system is composed of a perception module, a control and transmission module, and a motion module. It enables automated data collection at any position beneath the bridge structure. (2) A block-based panoramic generation strategy is employed, which uses a spatially ordered block concept to simplify the panorama stitching process and enhance accuracy. (3) Deep learning-driven two-phase synchronous identification and localization method. In the first phase, MobileNetV4 serves as the primary feature representation tool, facilitating the lightweight reconstruction of panoramic images. In the second phase, the YOLOv9 detection framework is employed to perform a precise analysis of the identified defect regions, providing detailed defect information on a localized level. The design of this system significantly enhances the efficiency and accuracy of inspections of large-span bridge undersides, offering robust technical support for bridge health maintenance. Experimental results indicate that the proposed method achieves over 90 % accuracy in defect recognition tasks, alongside millimeter-level precision in localization.