Existing deployed Unmanned Aerial Vehicles (UAVs) visual trackers are usually based on the correlation filter framework. Although these methods have certain advantages of low computational complexity, the tracking performance of small targets and fast motion scenarios is not satisfactory. In this paper, we present a novel multi-level prediction Siamese network (MLPS) for object tracking in UAV videos, which consists of Siamese feature extraction module and multi-level prediction module. The multi-level prediction module can make full use of the characteristics of each layer features to achieve robust evaluation of targets with different scales. Meanwhile, for small-size target tracking, we design a residual feature fusion block, which is used to constrain the low-level feature representation by using high-level abstract semantics, and obtain the improvement of the tracker's ability to distinguish scene details. In addition, we propose a layer attention fusion block which is sensitive to the informative features of each layers to achieve adaptive fusion of different levels of correlation responses by dynamically balancing the multi-layer features. Sufficient experiments on several UAV tracking benchmarks demonstrate that MLPS achieves state-of-the-art performance and runs at a speed over 97 FPS.

现有的已部署的无人机视觉跟踪器通常基于相关性过滤器框架。尽管这些方法具有计算复杂度低的某些优点，但小目标和快速运动场景的跟踪性能并不令人满意。在本文中，我们提出了一种用于无人机视频中目标跟踪的新型多级预测暹罗网络（MLPS），该网络由暹罗特征提取模块和多级预测模块组成。多级预测模块可以充分利用每个图层特征的特征，以实现对不同尺度目标的鲁棒评估。同时，对于小型目标跟踪，我们设计了残差特征融合块，用于通过使用高级抽象语义来约束低级特征表示，并提高跟踪器区分场景细节的能力。此外，我们提出了一个对每一层信息特征敏感的层注意融合块，以通过动态平衡多层特征来实现不同级别的相关响应的自适应融合。在几个无人机跟踪基准上进行的充分实验表明，MLPS达到了最先进的性能，并以超过97 FPS的速度运行。