Mining Urban Traffic Events and Anomalies

As huge numbers of sensors relay real-time traffic measurements on city roads, real-time information management systems need to be in place that sieve relevant from irrelevant information. This is particularly important in mission critical applications. For example traffic operators need to be informed about potential anomalies caused by road traffic accidents or similar events in order to deploy timely mitigation strategies. We investigate a new method that elevates anomalies visually to an operator by using a combination of change-detection analytics and auxiliary information from human-sensors (e.g., through Twitter). We identify operational regimes of constant behaviour and change points that bound a regime. This allows us to account for drifts that are not captured in a model. This has the advantage over cumulative sum (CUSUM) algorithms that no threshold parameter needs to be specified upfront and relaxes the assumption of stationarity. Stationarity assumptions do not hold in traffic systems, as network effects take hold and traffic signal control intervenes when local traffic states are deteriorating.