TrafXpert: Zero-Shot Road Anomaly Reasoning via Chain-of-Anomaly Thought

Jiaying Wu, Liqi Yan, Siqi Song, Yun Li

Abstract: With the rapid development of intelligent transportation systems (ITS), the complexity and dynamics of urban traffic continue to increase, making traffic anomaly detection and analysis more and more important. Traditional traffic anomaly detection methods have disadvantages such as poor flexibility, limited detection range, and overly strong assumptions about data distribution. Existing vision-language models also exhibit deficiencies in local feature sensitivity, causal reasoning, and zero-shot generalization when applied to road anomaly detection. To tackle these challenges, this study introduces TrafXpert, a zero-shot road anomaly reasoning framework that pushes the boundaries of generalization and reliability in traffic anomaly detection. First, this framework innovatively combines image features extracted by a visual encoder with multi-source data to construct a conditional probability model, and realizes the preliminary judgment of traffic scene anomalies by calculating the posterior probability. Second, it uses large models to mark seed samples and then generate a large number of question-answering data in batches, and generates high-quality enhanced corpus by designing specific Prompts to standardize the output. Third, based on a specific model, it goes through stages of perception, judgment, and classification to achieve accurate identification and classification of traffic anomalies. This framework can quickly and accurately detect various types of traffic anomalies, providing a new solution for anomaly detection in the field of intelligent transportation. All models and datasets will be publicly available.

External IDs:dblp:conf/prcv/WuYSL25