Road Entity Detection on the Austrian Highway Roadwork Views Dataset

Source abstract

Automated perception of road construction sites is essential for reliable monitoring and safe navigation on dynamically changing road infrastructure. However, developing robust computer vision methods for this domain requires representative and well-annotated data. In this paper, we present the Austrian Highway Roadwork Views (AHRV), a comprehensive annotated dataset of roadwork scenes recorded on Austrian highways. The dataset contains 1,812 scene images with 17,436 annotations of 7 object classes, focusing on temporary roadwork infrastructure such as barriers, traffic cones, traffic beacons, and signage. AHRV captures the considerable visual and structural variability of real-world construction sites, including changing lane configurations, diverse highway environments, and different acquisition settings. As a reference baseline, an RF-DETR Large model achieves an instance-weighted mAP50 of 93.5% on the complete test split. AHRV provides a dedicated benchmark for developing and evaluating object detection methods for highway construction sites, and the dataset, baseline model, and supporting code are made publicly available for future research.