Traffic-IT: Enhancing traffic scene understanding for multimodal large language models

In recent years, the convergence of artificial intelligence and urban infrastructure has driven transformative advances in intelligent transportation systems (ITS). However, traditional models often lack the generalizability needed to adapt to diverse traffic scenarios. Multimodal large language models (MLLMs) offer a promising solution, yet they are typically trained on general datasets, limiting their effectiveness in specific transportation contexts. To address this, we introduce Traffic-IT, a dataset comprising 220,950 question-and-answer pairs from 30,000 images, designed to enhance MLLMs’ capabilities in traffic scene understanding. The dataset covers various traffic scenarios, including weather conditions, locations, and times of day, providing in-depth insights and driving strategies tailored to real-world needs. Created through expert consultation and rigorous validation, Traffic-IT significantly improves MLLMs’ performance in interpreting complex traffic scenes. We anticipate that Traffic-IT will be a crucial resource for future developments in smart city applications.