This thesis focuses on the development of a robotic platform for bridge bearing inspection. The existing literature on this topic highlights an aspiration for increased automation of bridge inspection, due to an increasing amount of ageing infrastructure and costly inspection.
Furthermore, bridge bearings are highlighted as being one of the most costly components of the bridge to maintain.
However, although autonomous robotic inspection is often stated as an aspiration, the existing literature for robotic bridge inspection often neglects to include the requirement of autonomous navigation. To achieve autonomous inspection, some methods for mapping and
localising in the bridge structure are required. This thesis compares existing methods for simultaneous localisation and mapping (SLAM) with localisation-only methods. In addition, a method for using pre-existing data to create maps for localisation is proposed.
A robotic platform was developed and these methods for localisation and mapping were then compared in a laboratory environment and then in a real bridge environment. The errors in the bridge environment are greater than in the laboratory environment, but remained within a defined error bound. A combined approach is suggested as an appropriate method for combining the lower errors of a SLAM approach with the advantages of a localisation approach for defining existing goals. Longer-term testing in a real bridge environment is still required.
The use of existing inspection data is then extended to the creation of a simulation environment, with the goal of creating a methodology for testing different configurations of
bridges or robots in a more realistic environment than laboratory testing, or other existing simulation environments.
Finally, the inspection of the structure surrounding the bridge bearing is considered, with a particular focus on the detection and segmentation of cracks in concrete. A deep learning approach is used to segment cracks from an existing dataset and compared to
an existing machine learning approach, with the deep-learning approach achieving a higher performance using a pixel-based evaluation. Other evaluation methods were also compared
that take the structure of the crack, and other related datasets, into account.
The generalisation of the approach for crack segmentation is evaluated by comparing
the results of the trained on different datasets. Finally, recommendations for improving the
datasets to allow better comparisons in future work is given.
