Implementation of the Hungarian Method for object tracking on a camera monitored transportation system

In vision data processing often the positions of detected objects have to be determined. This can be done by many sophisticated algorithms using several characteristic optical properties of each object. The position of each object can then be fed into additional state estimation algorithms. Thus it is possible to use multiple successive frames to calculate a precise estimation of the state of each object containing position, orientation, velocity and its trajectory. If however many objects are detected but all objects look the same (no characteristic optical properties can be determined), tracking of objects becomes more difficult. The main problem that arises is to assign the detected object positions to the predicted position estimations which can be calculated from the state estimation for each object. In literature, this topic is referred to as multi target tracking. This paper describes a simple solution for a multi target tracking problem that arises, if multiple small scale transportation vehicles are operated in a camera monitored area. While the state estimation for each vehicle and obstacle is performed by a Kalman filter algorithm, the assignment of the detected object positions to the position predictions of all objects is done by the Hungarian method with some extension to avoid assignments of objects that are located too far apart from each other.