Zhu, Xiaolei (National University of Singapore), Bai, Qiaowen (National University of Singapore), Ong, Ghim Ping (National University of Singapore), Sikdar, Biplab (National University of Singapore)

Real-Time Traffic Incident Detection with Sparse Observations: A Masked Spatiotemporal Graph Learning Framework

Scheduled for presentation during the Regular Session "S32c-AI-Driven Traffic Monitoring, Safety, and Anomaly Detection" (FR-LA-T32), Friday, November 21, 2025, 16:40−17:00, Southport 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Timely and accurate incident detection is essential for maintaining the efficiency and resilience of transportation systems, especially under sparse sensing conditions. This paper proposes a Masked Spatiotemporal Graph Learning (MSTGL) framework for robust incident detection using sparse traffic observations. The framework integrates a spatial graph attention encoder to model localized traffic interactions and a BERT-based temporal encoder to capture evolving temporal dynamics. A masked supervision strategy is introduced during training to enhance the predictive capability of the traffic state decoder, which forecasts future traffic conditions from spatiotemporal embeddings. On top of this, we propose a dual-window anomaly detection module, which utilizes both short-term and long-term prediction residuals to identify incident-induced deviations in real time. The proposed method is evaluated on a campus-scale traffic network in Singapore with sparse fixed-point camera coverage. Experimental results demonstrate that the proposed framework significantly improves detection performance, achieving a 17.76% higher F1-score compared to state-of-the-art baselines.