Deep Learning for Road Infrastructure Monitoring: An Experimental Benchmark of Detection and Segmentation Models

We systematically evaluate the application of deep learning methods to road infrastructure monitoring, a crucial area for ensuring safety and sustainability in transportation systems. Specifically, our study delivers an experimental comparative benchmark evaluating the localization of road surface anomalies (e.g., road cracks, potholes, and faded markings) and infrastructural points of interest (e.g., traffic cones and signs). By leveraging heterogeneous datasets acquired from both aerial drones and ground-based inspection platforms, we rigorously evaluate the performance of various state-of-the-art computer vision models across object detection and semantic segmentation tasks. Our key contributions include a comprehensive comparative analysis of twelve object detection and seven semantic segmentation architectures across thirteen publicly available datasets, an assessment of their suitability for varying monitoring scenarios, and in-depth insights into how localization accuracy is shaped by the acquisition platform (inspection drone versus ground vehicle), defect type, and dataset-specific characteristics. The findings provide comparative baseline evidence regarding the performance of different architectures across specific road-monitoring tasks and dataset conditions, as well as underscoring the potential of deep learning-driven automation to significantly enhance road inspection processes, minimize operational costs, and facilitate proactive maintenance strategies. Thereby, this work establishes a robust foundation for future advancements in intelligent transportation systems and large-scale infrastructure management.