Traffic landmarks which we find painted on the roads to delimit road lanes, crosswalks, and speed limits are essential to ensure a smooth flow of traffic and the safety of pedestrians and drivers. One of the problems facing this aspect of the urban infrastructure is the deterioration of these landmarks due to prolongated periods of exposure to friction with car tiers and varying weather conditions, especially in rainy places such as Japan.
In Mie prefecture particularly, the sudden stop rate during road crossings reached 3.4% in the year 2020, which is indicative of a pressing problem relating to road safety. It has been deemed necessary by the local authorities of Mie prefecture to implement a reliable maintenance process for these road signs. Currently, the process of maintenance is inefficient and highly expensive. In this project, we try to implement an automatic camera-based monitoring system that can detect road signs from an image input and evaluate the level of deterioration using machine learning and deep learning techniques.
The development of this monitoring system will help the local government facilities of Mie prefecture in the detection of the areas of the roads in which the road signs are damaged and would also allow them to assess the degree of the deterioration and therefore allocate resources for repairing these damages more efficiently.
The monitoring system would consist of machine learning models operating on a set of servers. These servers would receive images from drive recorders of police cars alongside the GPS coordinates for each image. The monitoring module would then process these images in two steps:
- Traffic landmark detection
- Traffic landmark evaluation
The detection of traffic landmarks is achieved using a semantic segmentation model trained for this particular task.
Semantic segmentation models are a type of Convolutional Neural Networks which are used to detect objects in a given image with very high precision, such that each pixel in the target object is detected and separated from all the other objects in the image. The evaluation of Traffic landmarks is also accomplished by training a set of machine learning models of different types to obtain a score that indicates the quality of the detected traffic landmarks. Currently, the Information processing laboratory at Mie university is conducting research on state-of-the-art models, and the methods by which we can optimize their performance and apply them to the case study of this project in particular.