A Framework for Incident Detection in Large-Scale Public Bus Systems
Providing efficient urban mobility through quality public transportation is one of the main goals for using smart city technologies by public administrations. Disruptions caused by traffic incidents are a significant cause of delays in public bus systems. The literature shows several Incident detection (ID) Systems using multiple sources of information, such as camera images, radar (fixed sensors), and vehicles or smartphones with Global Positioning System (mobile sensors). Nevertheless, proposed ID Systems using mobile sensors mainly use cars and simulated car GPS data, while few works use bus GPS data.
This work presents a bus system’s ID model using bus GPS data on urban networks. We used historical bus GPS data to extract the characteristics of each segment of the bus system. We proposed two models for ID: a statistical approach based on the travel time of buses in each segment and a machine learning approach applied to multiple features describing real-time bus dynamics in each segment. To implement these models, we also proposed a framework architecture that enables the integration of scalable and efficient composite models. We implemented the framework and evaluated its scalability and the predictions of simple travel time models. We showed that using this framework can be feasible in real-time in large metropolitan areas, such as São Paulo city. Last, we implemented and evaluated the proposed ID models.
We defined traffic incidents based on changes in the headways between buses. We showed that our models achieved a detection rate of over 80% for long-duration (over 10 minutes) incidents with a mean time to detect less than 6.5 minutes for a false alarm rate of 3%.