结合全局光流的孪生区域提名网络目标跟踪算法

Combined Global Optical Flow with Siamese Region Proposal Network for Object Tracking

摘要:

基于孪生网络的跟踪器由于其均衡的跟踪速度和精度而成为目标跟踪领域的主流的研究框架。近年来，涌现出许多基于孪生网络的目标跟踪算法。然而，受限于孪生网络跟踪框架固有的跟踪机制和搜索区域选择机制，如何解决当目标处在被遮挡、快速运动和出视野等困难场景下时，依然稳定、鲁棒的进行目标跟踪始终是孪生网络跟踪器亟需解决的问题。为此，本文提出一种结合光流的孪生区域提名网络目标跟踪算法（GOF-SiamRPN）。通过全局光流对目标的运动趋势信息进行补充，本文提出的方法可以有效地解决这些困难场景下的跟踪问题。在VOT2019和UAV123上的实验结果表明，相比我们的基准方法，本算法分别取得了2.0%和1.8%的性能提升。与其他先进的跟踪器相比，本算法也取得了有竞争力的跟踪效果。这充分说明了本算法的有效性。

Abstract:

Siamese network-based trackers have become a mainstream research framework in object tracking due to their balanced tracking speed and accuracy. Many object-tracking algorithms based on Siamese networks have emerged in recent years. However, limited by the inherent tracking mechanism and search area selection mechanism of the Siamese network-based tracking framework, how to solve the problem of stable and robust object tracking when the object is under challenging scenarios such as occlusion, fast motion, and out-of-view is always an urgent problem for Siamese trackers. To this end, in this paper, we propose an object-tracking algorithm that combined the Siamese region proposal network with the global optical flow (GOF-SiamRPN). By assisting the motion trend information of the object with global optical flow, the proposed method can effectively solve the potential tracking issues in these difficult scenarios. Extensive experimental results on VOT2019 and UAV123 show that the proposed method achieves a performance gain of 2.0% and 1.8% compared with the baseline method, also achieves a competing performance compared to other state-of-the-art trackers, which fully validates its effectiveness.