Detecting Convoys Using License Plate Recognition Data

License plate recognition
(LPR) sensors are embedded camera systems that monitor road traffic. When a vehicle passes by a sensor, the vehicle's license plate, the location, and the time of observation are recorded. Given a stream of such observations from a collection of sensors spread around the road network, our goal is to detect convoys: groups of two or more vehicles traveling with highly correlated trajectories. Some of the main challenges with modeling and processing data from LPR sensors include that the data-gathering process is event-driven, thus data are not regularly sampled in time or space. Also, an appropriate definition of convoy should be relative to background traffic patterns, which are temporally and spatially varying. This paper proposes novel models for LPR observations of traffic which are well suited for online convoy detection. Baseline traffic is modeled as following a mixture of semi-Markov processes, and specific models for temporal and spatial correlation of observations of vehicles traveling in a convoy are introduced. These models are used within a sequential hypothesis testing framework to obtain a system for real-time convoy detection. The model of baseline traffic may be of independent interest for forecasting road traffic patterns. Experiments with an extensive simulated dataset illustrate the performance of the scheme and offer insights into the tradeoffs between detection rate, false alarm rate, and the expected number of observations required to detect a convoy.