In the thesis the problem of long-term visual tracking is addressed. The main challenges of the problem are on-line learning of the target's visual appearance, recognition of target's absence and it's redetection. A part-based tracker is proposed using local features and affine transformation. Long-term tracking is performed with tracking-by-detection, supported by optical flow in the short term. Two nested methods are used when fitting the transformation: firstly, a cluster of potential target points is defined, then the affine deformation is robustly estimated. New model features are added based on the global shape template, that is updated by the features themselves, forming a feedback-loop. The tracker is tested on two groups of sequences, the first targeting long-term and the second short-term trackers. The results are compared with the state-of-the-art methods. The performance of the tracker is comparable, though the problem of redetection should be more carefully addressed.