Infrastructure maintenance and management in Japan, which is said to be an "advanced" nation, is a serious problem due to the lack of financial resources and experts. On the other hand, in recent years, sophisticated image recognition has become possible because of the development of technologies such as deep learning, and the increased refinement of camera functions of smartphones, which are widely used worldwide.
In this research, in cooperation with road managers of seven municipalities, we conducted real-time detection of road surface damage by deep learning, and attempted to extract criteria for repair correspondence decision for each local government by random forest method. As a result, it was possible to detect road surface damage with a detection rate of 87% using only general smartphones, and it was possible to catch a glimpse of the difference in standards for maintenance level correspondence between local governments. With this result, inexpensive and simple infrastructure inspection becomes possible, and there is a possibility of breakthrough in various areas suffering from financial resources and expert shortage.
View full abstract