POTHOLE AND CRACK DETECTION USING DEEP LEARNING: ADVANCEMENTS IN ROAD SURFACE ANOMALY RECOGNITION

Abstract

Abstract: Problems with asphalt roads are a big problem for both developed and developing countries because they make it hard for people to get to and from work every day. Because they are so dangerous for drivers and riders, potholes are a common type of damage to pavement that has been studied a lot. The purpose of this research is to find out how well three advanced deep learning models work when used on external devices to find potholes. The point of our research was to find potholes using an edge platform made up of a Raspberry Pi single-board computer and an OAK-D AI module. This essay gives a complete assessment of how well different advanced object recognition frameworks and deep learning models work in real time at finding potholes. The project uses live video footage from a moving car and a collection of pictures showing potholes in different types of roads and lighting. The average precisions (mAPs) of 80.04% for Tiny-YOLOv4, 85.48% for YOLOv4, and 95% for YOLOv5 show that the suggested method is good for finding faults and can be used for real-time identification with OAK-D. It was decided that Tiny-YOLOv4 is the best model for real-time pothole recognition after the research was done at a frame rate of 31.76 FPS and a detection accuracy of 90%.Through the use of deep learning, and more specifically ResNet50 models, the process of locating cracks and potholes in road surfaces has become significantly simpler in recent years.