Tracking multiple vehicles using foreground, background and motion models

In this paper a vehicle tracking algorithm is presented based on the combination of a per pixel background model (an extension of work by Stauffer and Grimson [12]) and a set of single hypothesis foreground models based on a general model of object size, position, velocity, and colour distribution. Each pixel in the scene is thus ‘explained’ as either background, belonging to one of the foreground objects or as noise. Calibrated ground-plane information is used within the foreground model to strengthen the object size and velocity consistency assumptions. A learned a priori model of typical road travel direction and speed is used to provide a prior estimate of object velocity which used to initialise the velocity model of each of the foreground object models. This model is typically an Extended Kalman filter but other models are possible within the algorithm. The system runs at near video frame rate ( 20fps) on modest hardware and is robust assuming sufficient image resolution is available and vehicle sizes do not greatly exceed the priors on object size used in object initialisation.