A hybrid model approach for real-time visual object tracking

Tracking an unknown general object in video sequences is a challenging task in many computer vision applications since the appearance of the object can change significantly due to pose variations, illumination changes, shape deformations, and abrupt motions. In this dissertation, we address these object tracking challenges by building a hybrid model tracker which consists of a tracking by detection framework and a Kalman filtering framework.;We first present two tracking by detection approaches by applying the naive Bayes classifier and the support vector machine approach respectively. In both approaches, effective feature selection schemes are proposed to select the most informative features to construct a robust appearance model.;Then, we propose a novel approach for appearance estimation in object tracking. Most existing tracking algorithms assume that the object appearance is static for two consecutive frames, which remains a potential risk causing the drift problem. We build a Kalman filtering framework to generate a statistically optimal estimation for the object appearance. The proposed method greatly reduces the error between the true object appearance and the estimated object appearance, thus effectively improving the tracking performance.;Finally, we develop and implement a hybrid model tracking system which combines the discriminative model constructed in the tracking by detection framework and the generative model estimated by the Kalman filtering framework. The support vector machine tracker is applied to provide accurate feedback to the Kalman filtering framework, which improves the estimating precision. The Kalman filtering framework then generates the optimal estimation of the object appearance and determines the object location by the best fitting with the appearance model. The object location determined by maximizing the classifier confidence in the support vector tracking framework is finally corrected by the Kalman filtering framework. Therefore, the proposed tracking system provides more meaningful tracking results compared with traditional tracking by detection algorithms which suffer from the inconsistent objectives between tracking and classification. Our experimental evaluations show that a significant improvement over state-of-the-art methods is achieved by our approach.