Discriminative model-based trackers have made remarkable progress recently. However, due to the extreme imbalance of foreground and background samples, the learned model is hard to fit the training samples well in the online tracking. In this paper, to alleviate the negative influence caused by the imbalance issue, we propose a novel construction scheme of target-aware features for online discriminative tracking. Specifically, we design a sub-network to generate target-aware feature embeddings of foregrounds and backgrounds by projecting the learned feature embeddings into the target-aware feature space. Then, a model
solver, which is integrated into our networks, is applied to learn the discriminative model. Based on such feature construction, the learned model is able to fit training samples well in the online tracking. Experimental results on four benchmarks, OTB-2015, VOT-2018, NfS, and GOT-10k, show that the proposed target-aware feature construction is effective for visual tracking, leading to the high-performance of our tracker.