As original compressive tracking algorithms can not select adaptively the object futures, it will result in drifting or tracking lost when the object is occluded or its appearance changes. To address this problem, this paper proposes a real-time compressive tracking algorithm based on online feature selection. First, two complementary projection matrixes were generated in an initial phase, and the projection matrixes were used to extract the feature to construct a feature pool. Then, features with high confidence scores were selected from the feature pool by a confidence evaluation strategy and these discriminating features and their corresponding confidence scores were utilized to construct a Naive Bayesian classifier. Finally, the classifier was taken to process candidate samples by binary classification and response results to the classifier were taken as tracking results. However, the previous result was used to online update the feature pool and the classifier to prepare for subsequent processing. The tracking performance of proposed algorithm is compared with that of original compressive tracking algorithms on several public testing video sequences. The comparison shows that the proposed algorithm improves the tracking accuracy and robustness, and the processing frame rate is 25 frame/s on a 320 pixel×240 pixel video sequence, which meets the requirements of real-time tracking.