Extreme Traffic Monitoring using Probabilistic Inference
John Quinn
Makerere University, Kampala/Uganda


Computer vision work on traffic monitoring usually assumes a constrained environment such as a highway, where cars move uniformly within separate lanes. In some situations, for example in many cities in the developing world, extremely chaotic traffic conditions make such assumptions unrealistic. Traffic monitoring is usually done in two stages: estimation of road geometry and tracking of vehicles. We can make both these processes robust by incoporating our uncertainty into a probabilistic dynamical model.
In this talk I will show how we have produced robustified traffic monitoring by applying a dynamical model which uses optical flow and scale-invariant features to distinguish vehicles from clutter. Instead of tracking individual vehicles we characterise a lane of traffic as a fluid and quantify the flow rate and velocity.

back to the list of talks