ROADSIGHT: A novel dataset for real-time intersection detection in aerial scenes under seasonal variation

Akash Shingha Bappy, Siva Ariram, Timo Mäenpää, Juha Röning

Keywords: Road intersection detection, aerial imagery, UAV dataset, YOLOv11, edge computing, seasonal variation, real-time detection, deep learning

Abstract: Road intersections serve as critical nodes in transportation networks, and their accurate detection from aerial imagery is helpful for applications such as autonomous navigation, urban planning, and traffic management. However, existing datasets for object detection in aerial views often lack specificity to intersections, diversity in seasonal conditions, and customized annotations for unmanned aerial vehicle (UAV) captured data, which leads to challenges in model robustness and real-time performance on edge devices. This work introduces a novel dataset named ROADSIGHT (Road-Oriented Aerial DataSet for Intersection detection on edGe Hardware plaTform), a UAV-captured dataset of high-resolution RGB images collected in both summer and winter conditions, with expert bounding-box annotations for two classes: roundabouts and intersections (encompassing 3-leg T/Y and 4-leg types). We benchmark state-of-the-art models(YOLOv8, YOLOv11, YOLOv12, and RT-DETR) and identify YOLOv11s as edge suitable. Robustness is validated through cross-validation with stable mAP. This work contributes a focused, seasonally varying benchmark and corresponding baseline results for real-time intersection detection on resource constrained UAV platforms, while clearly acknowledging current limitations in geographic coverage, adverse weather, and class granularity for future extensions.

Primary Area: datasets and benchmarks

Submission Number: 11258