Review of Kalman Filters in Multiple Object Tracking Algorithms

In multiple object tracking algorithms, the Kalman filter is used as a predictor of the new position of each tracked object. This prediction is then used as one of the factors for matching new detections to already tracked objects. As the Kalman filter operates on an assumed motion model, the quality of its prediction depends on how accurate those assumptions are with respect to the actual movement of the objects as seen by the camera. To better understand and evaluate the accuracy of different Kalman filter implementations, previous works and those generated from noise models were compared. From our simulations, we have seen similar performance among all the tested Kalman filters when tuned, showing that simpler filters are more computationally efficient.