Abstract: Siamese region proposal network tracing algorithm is an efficient target tracing algorithm, where anchor box is used to avoid the influence of image pyramid to tracing performance. However, this tracing method is subject to limitation of region proposal network itself. While the target is rotating, the tracing accuracy is greatly reduced. As to other algorithms which demonstrate high rotation robustness, the adoption of complex rotation structure results in substantial decline of tracing speed. In order to address the influence of rotating target to the region proposal network tracing accuracy, a siamese network tracing algorithm of rotating region proposal network is proposed. Through unification of rotation and region proposal network via arbitrary-oriented region proposal network（AO-RPN） and by introduction of angle prediction branch, the rotating target is searched directly during target tracing, obtaining the minimum bounding rectangle. This method maintains high tracing speed and achieves higher accuracy than an algorithm where target is subject to rotating sampling or traced using local characteristics. The experimental results of OTB2015, VOT2016 and VOT2018 data sets suggest that the proposed algorithm demonstrates strong robustness and adaptability to multiple complex conditions such as blocking, deformation and lighting.

Key words: object tracing, feature fusion, rotating anchor box, region proposal network, siamese network

摘要： 孪生区域提议网络跟踪算法是一种高效的目标跟踪算法，通过锚框规避了图像金字塔对跟踪性能带来的影响，但这种跟踪方法受制于区域提议网络本身的局限性，在目标旋转时，跟踪精度将受到较大损失。而其他对旋转鲁棒性较高的方法则因为使用了复杂的旋转结构，导致算法的跟踪速度大幅下降。为了解决旋转目标对区域提议网络跟踪精度的影响，提出了旋转区域提议网络的孪生神经网络跟踪算法，通过AO-RPN（arbitrary-oriented region proposal network）结构将旋转与区域提议网络相统一，引入角度预测分支，在目标跟踪的过程中，直接对旋转的目标进行搜索，并得到最小外接矩形。该方法在保持较高跟踪速度的同时，精度超过了对目标进行旋转采样或使用局部特征进行跟踪的算法。通过在数据集OTB2015、VOT2016和VOT2018上进行的大量实验。结果表明，该算法在遮挡、形变、光照等多种复杂情况下表现出了较强的鲁棒性和适应性。

关键词: 目标跟踪, 特征融合, 旋转锚框, 区域提议网络, 孪生神经网络