Vision-Based Traffic Conflict Detection of Signalized Intersections

Sharma, Anuj · 2018 · ROSA P / Midwest Transportation Center

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 study addresses the challenge of automating the interpretation of driving environments using image processing, specifically within dynamically changing, naturalistic real-world settings. Current vision-based systems often struggle to simultaneously recognize multiple objects of interest, such as traffic signs, vehicles, and signals, which is necessary for understanding pre-crash causal factors like distracted or aggressive driving. The research was motivated by the need to maximize the utility of large-scale naturalistic driving datasets, such as those from the Second Strategic Highway Research Program (SHRP2), which contain low-resolution videos recorded under varying congestion, lighting, and weather conditions. The primary goal was to develop a robust vision system capable of capturing driver dynamics and detecting objects relevant to traffic safety. To achieve this, the researchers developed and tested a novel detector named Shape-ACF, which combines aggregate channel features with a shape descriptor. The study utilized data from the SHRP2 naturalistic driving study, creating a new annotated dataset comprising 17 object classes grouped into traffic information signs, speed limit signs, vehicles/tail brake lights, and traffic signal status. The test set included approximately two hours of footage from 37 distinct trips across signalized and non-signalized routes, intentionally covering diverse conditions including different times of day, weather variations, and video resolutions. The methodology involved comparing the Shape-ACF detector against a pure ACF-based detector and an R-CNN based detector, evaluating performance in terms of accuracy, processing time, and true positive rates. The results indicated that the precision of the vision system varied by object category. Traffic lights, cars, and vehicle brake lights were detected with approximately 85% precision, while traffic signs and speed limits achieved 80% and 70% precision, respectively. Specific challenges were identified: "Pedestrian Crossing" and "Signal Ahead" signs had high false positive rates due to confusion with similarly structured yellow signs, particularly in poor weather. Additionally, distinguishing between 25 mph and 35 mph speed limit signs was difficult in low-resolution or bad weather conditions, with about 25% of 25 mph detections incorrectly identified as 35 mph. Nighttime detection for critical safety objects dropped by roughly 20%, although brake lights were more easily detected in darkness. Overall, the Shape-ACF system demonstrated the highest true positive rate with lower false positives and faster processing speeds compared to the other detectors, making it particularly suitable for mobile devices with limited processing power, whereas R-CNN remained superior for high-accuracy tasks on specialized hardware like GPUs. The significance of this work lies in its potential to improve the understanding of baseline driving behaviors and identify risk factors contributing to hazardous situations, thereby aiding in the development of safety countermeasures. The proposed system offers a viable method for automating object detection in naturalistic settings, providing data currently unavailable in standard roadway databases. However, the study notes a gap in evaluating how detected objects influence driver behavior, such as relative distance and conflict prevalence. Future research is recommended to fuse vision system outputs with auxiliary data like GPS, laser, and accelerometer readings to provide a more comprehensive analysis of traffic conflicts and driver interactions.

Key finding

Shape-ACF detected traffic light status, cars, and brake lights at about 85% precision but speed limit signs at only about 70%, with precision falling roughly 20% at night.

Methodology

modeling

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 (7 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.

• hazard perception
• sign visibility legibility
• looked but failed to see
• distraction detection algorithms
• motorcycle conspicuity
• naturalistic crash near crash

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: dataset resource