A multiple model PHD approach to tracking of cars under an assumed rectangular shape

This paper presents an extended target tracking method for tracking cars in urban traffic using data from laser range sensors. Results are presented for three real world datasets that contain multiple cars, occlusions, and maneuver changes. The car's shape is approximated by a rectangle, and single track steering models are used for the target kinematics. A multiple model approach is taken for both the dynamics modeling and the measurement modeling. A comparison to ground truth shows that the estimation errors are generally very small: on average the absolute error is less than half a degree for the heading. Multiple cars are handled using a multiple model PHD filter, where a variable probability of detection is integrated to enable tracking of occluded cars.