The Path Inference Filter: Model-Based Low-Latency Map Matching of Probe Vehicle Data

We consider the problem of reconstructing vehicle trajectories from sparse sequences of GPS points, for which the sampling interval ranges between 10 seconds and 2 minutes. We introduce a new class of algorithms, collectively called path inference filter (PIF), that maps streaming GPS data in real-time, with a high throughput. We present an efficient Expectation Maximization algorithm to train the filter on new data without ground truth observations. The path inference filter is evaluated on a large San Francisco taxi dataset. It is deployed at an industrial scale inside the Mobile Millennium traffic information system, and is used to map fleets of vehicles in San Francisco, Sacramento, Stockholm and Porto.