Because of the poor efficiency and effectiveness of current visual tracking algorithms, a real-time object tracking algorithm is proposed based on subspace learning. Under the framework of particle filtering, this paper uses the incremental PCA subspace method to learn an orthogonal subspace, and then get the linear representation of target appearance. In order to avoid the tracking drift produced by complicated interference, such as occlusions, motion blur and so on, an observation model and a template update scheme are built, which consider the complicated interference especially occlusions, to solve the drift problem of the traditional observation model based on minimum mean square error. The experimental results show that the algorithm in complicated conditions can be well implemented compared with several state-of-the-art algorithms.