M. Divjak and D. Zazula (Slovenia)
Motion tracking, hand & face recognition, Kalman filter.
Paper presents a tracking algorithm especially suited for tracking of elastic objects whose appearance changes significantly from one video frame to another. The core of the method is a prediction algorithm comprised of four Kalman filters. The first one tracks object boundaries through video sequence, the second and the third improve the performance of the first one and the fourth Kalman filter smoothes the object's final motion trajectory. The obtained results show that our prediction algorithm accurately detects about 20 % more objects than the basic version of our face and hand tracker which uses no prediction. Resulting object regions are approximately 33 % more compact and about 31 % more accurate, while the accuracy of motion trajectories is reduced by 3 %.
Important Links:
Go Back
