Real-time vehicle tracking in image sequences

We present an algorithm for tracking vehicles through an image sequence. The algorithm is based on matching a model in the image plane to each observed vehicle. The model is based on the characteristic edges of an intensity image of a vehicle. This model is applied to track vehicles through image sequences. We introduce three refinements to a standard model based tracking approach. As a first refinement, we use Kalman filtering to control the position and scale of the models. A multiple hypotheses strategy is suggested to avoid mismatches to local edges with a similar structure as (a part of) the model. As a last refinement we dynamically adapt each model to the vehicle if is being matched to. The refined model is more characteristic for the tracked vehicle and therefore increases the accuracy and robustness of the track. Each of these refinements contributes in their own way to a better overall performance of the tracking algorithm. We show that vehicles can be tracked in real-time with off the shelf processing capabilities and that our method is capable to track objects in ambiguous situations. Experiments based on practical data are presented to underline these conclusions.