FARSA: Fully Automated Roadway Safety Assessment

Song, Weilian; Workman, Scott; Hadžić, Armin; Zhang, Xu; Green, Eric; Chen, Mei; Souleyrette, Reginald R.; Jacobs, Nathan · 2018 · OpenAlex-citations

DOI: 10.1109/wacv.2018.00063

archive: archived pipeline: cataloged verified

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

Summary

This paper introduces FARSA (Fully Automated Roadway Safety Assessment), a deep learning framework designed to automate the assessment of road safety using street-level panoramas. The research is motivated by the limitations of the US Road Assessment Program (usRAP), which assigns safety ratings from one to five stars based on manual, labor-intensive annotation of roadway features. This manual process is slow, costly, and prone to variability, hindering the ability of highway authorities to conduct systemic safety analyses across large regions. The authors propose a fully automated alternative that estimates star ratings and auxiliary roadway attributes in milliseconds, enabling rapid, large-scale safety assessments that could inform infrastructure investments and reduce fatalities. The method employs a deep convolutional neural network (CNN) architecture built upon a modified VGG-16 backbone. The network processes cropped, resized street-level panoramas to predict a categorical distribution over the five-tier star rating scale. To address the challenge of identifying small, safety-critical features in high-resolution imagery, the model incorporates task-specific spatial attention layers that allow the network to focus on relevant regions of the panorama for each prediction task. The training strategy combines three loss components: a supervised loss for the primary star rating task, a multi-task loss for estimating auxiliary attributes (such as curvature, median type, and roadside hazards), and an unsupervised loss based on Tobler’s First Law of Geography, which encourages similar predictions for geographically proximate panoramas. This semi-supervised approach expands the effective training set without requiring additional manual annotations. The system was evaluated on a dataset of 7,288 annotated panoramas from urban and rural regions in two US states, alongside approximately 36,000 pairs of unsupervised panorama pairs. An ablation study demonstrated that the full architecture, combining attention mechanisms, multi-task learning, and unsupervised regularization, achieved a top-1 accuracy of 46.91%, outperforming baseline models and variants lacking specific components. The multi-task learning component significantly improved performance on auxiliary tasks, with accuracy gains of up to 47% over random baselines for attributes like area type and lane width. Visualizations of the attention masks confirmed that the network correctly focused on relevant features, such as driver-side hazards or medians. Additionally, the authors demonstrated a practical application by integrating safety scores into a routing engine, showing that the system could generate routes that prioritize safety over speed, avoiding low-rated road segments. The significance of this work lies in its potential to democratize and scale roadway safety analysis. By automating the usRAP rating process, FARSA reduces the cost and time required for systemic safety assessments, allowing both small and large agencies to identify high-risk road segments more frequently. This capability supports data-driven infrastructure planning and could contribute to reducing highway crashes and fatalities. The study establishes that combining multi-task learning with semi-supervised techniques effectively mitigates overfitting in small datasets, providing a robust framework for automated urban perception tasks.

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.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

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

• telematics crash prediction