NARMAD: A Novel Aerial Road Dataset with Oriented Annotations for Granular Traffic Analysis

The advancement of Intelligent Transportation Systems (ITS) in complex, non-lane-based traffic environments is critically dependent on high-quality, detailed datasets. This paper introduces NARMAD (Novel Aerial Road Mapping and Annotation Dataset), a large-scale, high-resolution dataset created from aerial videos capturing diverse and unstructured vehicle traffic From Ahmedabad and Surat, Gujarat, India. Distinct from existing aerial datasets, NARMAD provides fine-grained Oriented Bounding Box (OBB) annotations for 7 vehicle classes. The dataset contains over 516 images and 79,547 annotated instances, presenting significant challenges such as severe occlusion, vast scale variation, and high object density. We provide a comprehensive statistical analysis of the dataset and establish strong baseline results using state-of-the-art YOLO-based Oriented Bounding Box (OBB) detectors with tracking. NARMAD aims to provide better vehicle detections in complex and unorgranized traffic conditions and extension of trajectory prediction, and nuanced traffic behavior analysis in real-world urban settings.