[论文解读] Towards Robust Visual Tracking for Unmanned Aerial Vehicle with Tri-Attentional Correlation Filters

Yujie He, Changhong Fu+3 more|arXiv (Cornell University)|Aug 2, 2020

Video Surveillance and Tracking Methods参考文献 41被引用 4

一句话总结

该论文提出了一种三注意力相关滤波器（TACF）跟踪器，通过将上下文、维度和时空注意力机制整合到相关滤波框架中，实现鲁棒的无人机视觉跟踪。该方法在保持约28 FPS实时性能的同时，增强了对部分遮挡和杂乱背景的判别能力，在UAVDT和UAV123@10fps基准上相对于基线KCC跟踪器分别实现了4.8%和8.2%的成功率相对提升。

ABSTRACT

Object tracking has been broadly applied in unmanned aerial vehicle (UAV) tasks in recent years. However, existing algorithms still face difficulties such as partial occlusion, clutter background, and other challenging visual factors. Inspired by the cutting-edge attention mechanisms, a novel object tracking framework is proposed to leverage multi-level visual attention. Three primary attention, i.e., contextual attention, dimensional attention, and spatiotemporal attention, are integrated into the training and detection stages of correlation filter-based tracking pipeline. Therefore, the proposed tracker is equipped with robust discriminative power against challenging factors while maintaining high operational efficiency in UAV scenarios. Quantitative and qualitative experiments on two well-known benchmarks with 173 challenging UAV video sequences demonstrate the effectiveness of the proposed framework. The proposed tracking algorithm favorably outperforms 12 state-of-the-art methods, yielding 4.8% relative gain in UAVDT and 8.2% relative gain in UAV123@10fps against the baseline tracker while operating at the speed of $\sim$ 28 frames per second.

研究动机与目标

• 解决基于相关滤波的跟踪器在无人机场景下，特别是在部分遮挡和背景杂乱情况下的局限性。
• 通过将多级视觉注意力整合到相关滤波的训练与检测流程中，提升判别能力。
• 在提升对复杂视觉条件的鲁棒性的同时，保持适合机载无人机系统使用的实时性能。
• 开发一个统一的注意力框架，利用上下文、维度和时空线索，提升响应图质量。

提出的方法

• 引入上下文注意力模块，在滤波器训练期间为上下文块动态分配惩罚因子，提升环境感知能力和目标判别能力。
• 应用维度注意力，聚焦于响应图中有信息量的特征维度，通过抑制不相关特征降低噪声影响。
• 采用时空注意力，利用连续帧之间的时序一致性增强响应图，提升定位精度。
• 将上下文、维度和时空三种注意力机制融合到统一的三注意力相关滤波（TACF）框架中，实现端到端跟踪。
• 利用傅里叶域运算保持计算效率，实现在单个CPU上约28 FPS的实时推理。
• 在维度注意力模块中引入可学习的激活阈值βt，经优化后在UAVDT上达到最佳性能，取值为0.1。

Figure 1: Comparison between baseline KCC tracker [ 4 ] and the proposed TACF tracker. In UAV tracking, TACF incorporates the contextual information into the filter training stage, so that the attentional filters can sense the varying appearance of the object and environment simultaneously. After th

实验结果

研究问题

• RQ1 能否将多级注意力机制整合到相关滤波器中，以提升无人机跟踪在部分遮挡下的鲁棒性？
• RQ2 上下文注意力如何在不引入过多噪声的前提下，通过引入环境上下文增强滤波器训练？
• RQ3 维度注意力和时空注意力在复杂无人机场景中，对响应图质量和跟踪精度的提升程度如何？
• RQ4 所提出的三注意力框架是否能在显著优于最先进跟踪器的准确率和鲁棒性的同时，保持实时性能？

主要发现

• 与基线KCC跟踪器相比，TACF跟踪器在UAVDT基准上实现了4.8%的成功率相对提升。
• 在UAV123@10fps基准上，TACF相比基线实现了8.2%的成功率相对提升，成功率达到0.456。
• 消融实验表明，每个注意力模块（CA、DA、SA）均独立贡献性能，完整TACF模型在UAVDT上达到最高成功率为0.437。
• 每帧平均处理时间为35.59 ms（28.1 FPS），表明即使增加了注意力模块，仍具备实时可行性。
• 维度注意力的最优激活阈值βt被确定为0.1，在UAVDT上实现峰值性能。
• TACF跟踪器在两个主要无人机基准上均优于12种最先进跟踪器，在准确率和鲁棒性方面表现更优。

Figure 2: The main workflow of the TACF tracker. The object and context patches near the object are first extracted and represented as training samples. Then, the contextual attention strategy is implemented to realize adaptive learning of the surrounding environment throughout the online tracking p