The Object tracking is a challenging problem in computer vision field. Now, deep learning has made outstanding achievements in feature extraction. There are already some examples of deep learning applications in visual tracking. But, they almost paid attention to precision and sacrificed efficiency. So we propose a tracker named adaptive combination tracker(ACT), based on neural network model, which gives consideration to both precision and efficiency. Our ACT tracker gets rich features offline through Stacked Denoising Autoencoders (SDAE) and transfers them to online tracking. In tracking, a dynamic particle filter algorithm is proposed to speed up tracking. We update our tracker with different patterns according to the change in appearance. In order to ensure efficiency, our tracker regulates speed adaptively. Our tracker regulates speed precisely to avoid drift caused by fast motion as much as possible, and confirm and optimize the effectiveness of the results by comparison with previous target. We evaluate our tracker on OTB2013 datasets. Extensive experiments on OTB2013 datasets demonstrate that the proposed tracker performs favorably in relation to the state-of-the-art methods.