RadNet: Incident Prediction in Spatio-Temporal Road Graph Networks Using Traffic Forecasting
This work addresses the problem of minimizing service downtime and optimizing performance in road traffic systems for transportation management, but it is incremental as it builds on existing spatio-temporal forecasting approaches.
The paper tackles incident prediction in road traffic systems by developing RadNet, a neural model that forecasts system parameters like average vehicle speeds and compares them to historical averages to detect anomalies such as accidents, achieving up to 8% higher F1 scores compared to state-of-the-art methods.
Efficient and accurate incident prediction in spatio-temporal systems is critical to minimize service downtime and optimize performance. This work aims to utilize historic data to predict and diagnose incidents using spatio-temporal forecasting. We consider the specific use case of road traffic systems where incidents take the form of anomalous events, such as accidents or broken-down vehicles. To tackle this, we develop a neural model, called RadNet, which forecasts system parameters such as average vehicle speeds for a future timestep. As such systems largely follow daily or weekly periodicity, we compare RadNet's predictions against historical averages to label incidents. Unlike prior work, RadNet infers spatial and temporal trends in both permutations, finally combining the dense representations before forecasting. This facilitates informed inference and more accurate incident detection. Experiments with two publicly available and a new road traffic dataset demonstrate that the proposed model gives up to 8% higher prediction F1 scores compared to the state-of-the-art methods.