Accurate Traffic State Prediction with Deep Learning - Analyzing Statistical Aides for Identifying Anomalous Traffic Trends

• Published in: Proceedings - International Conference on Parallel and Distributed Systems pp. 408 - 414
• Main Authors: Hancock, Jaxon, Ren, Jian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.10.2024
• Subjects: Adaptation models; Analytical models; Computational modeling; Data models; Long short term memory; LSTM; Machine learning; Neural networks; Photography; Satellite images; security; smart mapping; traffic state estimation; Web services
• ISSN: 2690-5965
• DOI: 10.1109/ICPADS63350.2024.00060

Traffic is an expected part of many people's lives. To be able to accurately estimate the traffic state and determine the optimal route is of crucial importance to maintain coherence in both a smart-city system and in long-distance travel scenarios. Applications such as Google Maps, Apple Maps and Waze are critical tools in the hands of everyday commuters and emergency service providers alike-but also a potential vulnerability for those seeking to influence or disrupt regular traffic patterns. In this paper, we consider the most basic scenario-where the majority of end-users of mapping software are benign and truthful. We go on to propose a lightweight and adaptable model for estimating traffic states and demonstrate its effectiveness. We also analyze the feasibility of the proposed model for identifying anomalous traffic attacks in comparison with similar models. We create a theoretical framework based on minimization of error rate and conduct simulations which support efficacy in identifying these attacks using a statistically-augmented lightweight LSTM network with up to 84% fewer parameters than existing models, promoting use in low-resource applications. Moreover, the proposed model relies only on data which would be readily available to a smart mapping application. In 4-hour estimation scenarios with real-world data, the proposed model performed 11.8% better than a comparably complex LSTM model without statistical augmentation.