A new subspace-based tracking algorithm is proposed to deal with rapid changes and severe occlusion of targets during target tracking. Distance invariants are used to select superior pairs of SIFT features. Affine transformation parameters of affine transformation are calculated via linear fitting using these superior pairs. PCA subspace representation of the target is calculated via quick iterations. An effective observation model is constructed to track targets by combining this representation with obtained parameters under the framework of particle filter tracking. And SIFT features and PCA subspace are updated via online learning regularly. At the end, lots of experiments are performed and it is certificated that the proposed algorithm is able to precisely track targets of which gestures and shapes vary rapidly.