In this paper, a object tracking method based on spatio-temporal discriminative correlation filter is proposed. Firstly, a correlation filter layer is added into the the Siamese fully convolutional network to achieve end-to-end learning representation; secondly, the semantic feature is combined with the appearance feature to further enhance the discriminative ability of Siamese fully convolutional network; finally, the spatio-temporal regularized correlation filter is utilized to reduce the training time and improve the tracking performance. Extensive experiments conducted on VOT2017 dataset demonstrate the superior performance of the proposed approach over the examined state-of-the-art approaches.