REALTIME ROAD DAMAGE DETECTION

Abstract

A real-time road damage detection system is introduced for mobile deployment using deep learning techniques. Two object detection models Faster R-CNN and YOLOv12m are trained and evaluated on the RDD2022 dataset to determine their suitability for practical application. While Faster R-CNN achieves higher accuracy, its slower inference speed limits responsiveness. YOLOv12m, enhanced with architectural improvements such as R-ELAN, Area Attention, and FlashAttention, offers a better trade-off between speed and accuracy. Based on evaluation results, YOLOv12m is selected and integrated into an Android system, the system integrates the YOLOv12m model into a backend server and connects to the Android application via WebSocket, enabling real-time display of detection results on the user interface.