A TWO-STAGE EDGE-SERVER FRAMEWORK FOR REAL-TIME ROAD DAMAGE DETECTION USING DEEP NEURAL NETWORKS

Abstract

Timely detection of road surface defects is a prerequisite for proactive road maintenance, yet manual visual inspection remains slow, costly and subjective. Deep neural networks make automated detection from vehicle-mounted cameras and smartphones feasible, but detector families differ substantially in computational cost, and the model that is most suitable for a server is rarely suitable for an edge device. This paper analyses the deployment trade-offs of three representative detectors, YOLOv5s (single-stage), Faster R-CNN (two-stage) and Mask R-CNN (instance segmentation), for road damage detection on the RDD2020 dataset. An analysis of the dataset (21,041 images, eight damage types) shows a strong class imbalance of about 186:1 between the most and the least frequent class. Published results under the protocol of the Global Road Damage Detection Challenge show that single YOLOv5 and Faster R-CNN models reach comparable F1-scores (0.47-0.54). Inference measurements on an Apple M1 Pro laptop show that YOLOv5s processes 36 FPS on the integrated GPU and is 3.8 and 11.5 times faster on CPU than Faster R-CNN and Mask R-CNN, respectively, while having about six times fewer parameters. Based on these findings, a two-stage edge-server framework is proposed in which a lightweight detector performs real-time screening on the edge device and a region-based model performs detailed assessment on a municipal server. A deployment decision matrix and practical recommendations for cities of Kazakhstan, such as Almaty, are presented.