Video-Based Traffic Anomaly Detection with Vision-Language Models: A Survey

Abstract

Traffic anomaly detection is a hot research topic in road safety. With the rapid growth of video data, Video-based Traffic Anomaly Detection (VTAD) becomes a core module in safe driving and the security of surveillance systems. Visual unimodal methods are usually limited to shallow modeling of visual features only and lack linguistic reference frames, facing inherent shortcomings such as limited semantic comprehension and insufficient cross-domain generalization. In recent years, the rapid development of Vision-Language Models (VLMs) provides a new paradigm for traffic anomaly detection, which significantly improves the detection robustness in complex scenes. This paper provides the first survey of traffic anomaly detection based on VLMs, focusing on the three dominant methodologies: prompt learning-based, end-to-end fine-tuning-based, and feature adapter-based. In addition, in order to support and promote further research in the field, we provide a critical review of the latest traffic anomaly datasets and related evaluation metrics. Through this survey, we hope to provide valuable references and open possible trends for researchers and practitioners in the field.