Convolutional neural networks (CNNs) have shown favorable performance in recent tracking benchmark datasets. Some methods extract different levels of features based on pre-trained CNNs to deal with various challenging scenarios. Despite demonstrated successes for visual tracking, utilizing features from the same network might suffer from the suboptimal performance due to limitations of CNN architecture itself. We observe that different CNNs usually have complementary characteristics in representing target objects. Therefore, we propose to leverage the complementary properties of different CNNs for visual tracking in this paper. The importances of different CNNs are identified by a joint inference of candidate location, predicted location and confidence score. The prediction scores of all CNNs are adaptively fused to obtain robust tracking performance. Moreover, we introduce the attention mechanism to highlight discriminative features in each CNN. Experimental results on OTB2013 and OTB2015 datasets show that the proposed method performs favorably compared with some state-of-the-art methods. We conclude that combination of complementary models can better track objects in terms of accuracy and robustness.

卷积神经网络（CNN）在最近的跟踪基准数据集中已经显示出良好的性能。一些方法基于预训练的CNN提取不同级别的特征，以应对各种挑战性场景。尽管在视觉跟踪方面取得了成功，但由于CNN架构本身的局限性，利用同一网络的功能可能会导致性能欠佳。我们观察到，不同的CNN通常在表示目标对象时具有互补的特征。因此，我们建议在本文中利用不同CNN的互补属性进行视觉跟踪。通过对候选位置，预测位置和置信度得分的联合推断，可以确定不同CNN的重要性。将所有CNN的预测分数进行自适应融合以获得鲁棒的跟踪性能。此外，我们介绍了注意机制以突出显示每个CNN中的区别特征。在OTB2013和OTB2015数据集上的实验结果表明，与某些最新方法相比，该方法具有良好的性能。我们得出结论，互补模型的组合可以在准确性和鲁棒性方面更好地跟踪对象。