CrashFormer: A Multimodal Architecture to Predict the Risk of Crash
Published in ACM SIGSPATIAL: Advances in Urban-AI, 2023
Recommended citation: Karimi Monsefi, A., Shiri, P., Mohammadshirazi, A., Karimi Monsefi, N., Davies, R., Moosavi, S., & Ramnath, R. (2023, November). CrashFormer: A Multimodal Architecture to Predict the Risk of Crash. In Proceedings of the 1st ACM SIGSPATIAL International Workshop on Advances in Urban-AI (pp. 42-51). https://dl.acm.org/doi/abs/10.1145/3615900.3628769
Summary
This paper introduces CrashFormer, a novel multi-modal architecture designed to predict traffic accidents. Addressing the limitations of previous studies in generalizability and practical application, CrashFormer integrates diverse data sources including historical accidents, weather, map imagery, and demographic information. Covering a geographic area of 5.161 square kilometers and updating every six hours, the model consists of five components: sequential and image encoders, a raw data encoder, a feature fusion module, and a classifier. Tested in 10 major US cities, CrashFormer demonstrates a 1.8% improvement in F1-score over existing models, even with sparse input data, enhancing its applicability in real-world traffic safety scenarios.
Authors
Amin Karimi Monsefi, Pouya Shiri, Ahmad Mohammadshirazi, Nastaran Karimi Monsefi, Ron Davies, Sobhan Moosavi, Rajiv Ramnath