TL;DR
This paper presents a new framework that moves beyond evaluating daytime or nighttime factors individually, proposing a method to comprehensively assess traffic sign condition by combining Vision Language Models (VLMs) and LiDAR for both daytime visual performance and nighttime retroreflectivity.
Problem
Traffic signs are crucial for road safety, but traditional manual inspections are subjective, labor-intensive, and pose safety concerns. Retroreflectometers are expensive and unaffordable for smaller agencies. Most existing studies focus on either daytime factors or nighttime retroreflectivity, rarely integrating both aspects comprehensively.
Approach
Daytime Assessment: Three Vision Language Models (LLaVA, Qwen, InternVL) are fine-tuned to assess four key factors: legibility, color, surface and shape integrity, and surrounding environment. Model predictions are converted to numerical scores via sentiment analysis and CLIP scoring.
Nighttime Assessment: Nighttime retroreflectivity is extracted from LiDAR data following established calibration procedures.
Integration & Indexing: Daytime and nighttime assessment results are integrated to derive a comprehensive 'Sign Condition Index (SCI)' for maintenance guidance.
Results & Contribution
LLaVA and Qwen outperformed InternVL, achieving bidirectional cosine similarity scores of 0.67-0.76 across all factors.
Among 462 validated traffic signs, 68 were flagged as requiring immediate replacement due to inadequate retroreflectivity performance.