Automatic Detection and Localization of Roadwork

Narasimhan, Srinivas; Tamburo, Robert; Mertz, Christoph; Reddy, Dinesh; Vuong, Khiem; Ghosh, Anurag; Srivastava, Shefali; Boloor, Neha; Ma, Tiffany; Cardei, Michael; Dunn, Nicholas; Zhu, Hailiang · 2023 · ROSA P / Mobility21, Carnegie Mellon University

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This research addresses the challenge of automatically detecting and localizing roadwork zones to improve vehicular mobility, safety, and navigation for autonomous and assisted driving systems. Roadwork causes significant congestion and poses risks to driver-assist technologies, which often fail to navigate these heterogeneous environments, particularly when lane markings are obscured or altered. The authors aim to develop computer vision methods using affordable camera data to identify work zones, enabling dynamic route planning and real-time warnings. The study involved creating a novel dataset of 8,556 images from 19 U.S. cities, manually annotated with 26 classes of roadwork objects (e.g., cones, barriers, workers) and environmental surfaces. To address data scarcity for rare objects, the researchers developed "GeometryPaste," a data augmentation technique that scales and pastes segmented objects into new backgrounds using geometric context and vanishing point estimation. For road marking detection, they employed a Mask2Former model, pre-trained on the Mapillary Vistas dataset and fine-tuned on the Roadbotics dataset using pseudo-labeling to transfer lane marking annotations. Finally, for work zone detection and localization, they implemented a dual-model pipeline: an EfficientNet-B3 classifier for binary detection and a Mask R-CNN instance segmentation model for localization, with outputs processed via a convex hull algorithm to define the work zone boundary. The findings demonstrate the successful creation of a comprehensive, publicly available roadwork dataset. The GeometryPaste augmentation method improved instance segmentation performance for underrepresented objects like TTC message boards by ensuring realistic scaling and context. The Mask2Former approach effectively transferred lane marking detection capabilities to the roadwork domain without manual annotation. The dual-model detection and localization system achieved robust performance, with the EfficientNet classifier accurately identifying work zones and the Mask R-CNN providing precise spatial localization. The convex hull processing helped mitigate false positives by encapsulating scattered work zone objects into a coherent region. The significance of this work lies in providing a foundational dataset and effective algorithms for understanding roadwork environments. By enabling accurate detection and localization, these methods can enhance the safety of autonomous vehicles and driver-assist systems, preventing incidents such as vehicles entering construction sites. Furthermore, the aggregated data could support dynamic traffic management and navigation systems, offering real-time updates on lane restrictions and active work zones to optimize traffic flow and reduce congestion.

Key finding

A dual-model pipeline combining EfficientNet for detection and Mask R-CNN with convex hull processing for localization, supported by a novel geometry-aware data augmentation method, effectively identifies and delineates roadwork zones from visual data.

Methodology

dataset

Sample size: 8556

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• emergency work zone conspicuity