传统特征和深度特征融合的红外空中目标跟踪

新型红外诱饵的出现,对传统红外成像空空导弹作战效能的发挥造成了严峻挑战。近年来深度学习研究进展迅速,有力促进了目标跟踪领域的发展。以多域学习网络框架为基础,引入传统特征长宽比和均值对比度,将深度特征与传统特征融合在一个跟踪框架中,解决了单一特征在目标跟踪中无法有效对抗面源等复杂干扰的问题。为了评估算法性能,分别在仿真序列和实拍图像序列上进行了测试。实验结果表明,所提出算法的跟踪精度和鲁棒性优于目前经典的跟踪算法,是一种具有较强适应性的红外空中目标跟踪方法。

Infrared aerial target tracking based on fusion of traditional feature and deep feature

The emergence of new infrared decoys poses a severe challenge to the operational effectiveness of conventional infrared imaging air-to-air missiles. In recent years, the research progress of deep learning is rapid, which strongly promotes the development of target tracking and detection. Based on the MDNet framework, the aspect ratio and mean contrast of the artificial features are introduced, and the deep feature and the artificial feature are fused into a tracking framework, which solves the problem that the single feature could not effectively resist the complex interference such as the surface-type decoy in target tracking. In order to evaluate the performance of the algorithm, the simulation sequences and the real shot sequences are tested respectively. Experimental indicates show that the proposed algorithm is better than state-of-the-art trackers in both tracking accuracy and robustness, which is a kind of infrared aerial target tracking method with a strong adaptability.