Abstract
At present, manhole covers seem to cover a large percentage of the road surface. If a manhole cover becomes worn, it must be replaced because a worn cover can cause skidding and accidents. Staff have to check the condition of all covers periodically, and replace them if they are damaged. Because it takes too much time and manpower to check them manually, a system that can check them automatically by some remote sensor is urgently required. However, the detection of manhole cover wear has not been well studied. Our solution is to use image processing. We propose example-based approach using LBP feature, which is useful for texture recognition, considering the wear estimation problem as texture recognition. Feature selection utilizing measured data of wear to use only helpful features for wear estimation and patch extraction from images to estimate wear locally realize accurate estimation in spite of large variety of environments. Experiments with actual photos of manhole covers show that our approach achieved wearing detection accuracy scores of 0.95 in ROC/AUC scores, which is sufficient to help actual maintenance.
