Fusing multiple 2D visual features for vehicle detection

This contribution presents a multisensor fusion approach for vehicle detection. Shadow and symmetry features, abstracted to 3D sensors by including street surface information, are combined in an interacting multiple model filter with two system models, one for constant velocity, one for constant acceleration. Measurements are associated to tracks by means of a cheap joint probabilistic data association technique. Virtual prediction steps are used to incorporate multiple sensors and to achieve a versatile fusion architecture that allows easy integration of further sensors