Visual object tracking is an essential enabler for the automation of UAVs. Recently, Siamese network based trackers have achieved excellent performance on offline benchmarks. The Siamese network based trackers usually use classic deep and wide networks, such as AlexNet, VggNet, and ResNet, to extract the features of template frame and detection frame. However, due to the poor computing power of embedded devices, these models without modification are too heavy on calculation to be deployed on UAVs. In this paper, we propose a guideline to design a slim backbone: the dimension of output should be smaller than that of the input for every layer. Directed by the guideline, we reduce the computational requirements of AlexNet by 59.4%, while the tracker maintains a comparable accuracy. In addition, we adopt an anchor-free network as the tracking head, which requires less calculation than that of anchor-based method. Based on such approaches, our tracker achieves an AUC of 60.9% on UAV123 data set and reaches 30 frames per second on NVIDIA Jetson TX2, which, therefore, can be embedded in UAVs. To the best of our knowledge, it is the first real-time Siamese tracker deployed on the embedded system of UAVs. The code is available at GitHub.

视觉对象跟踪是无人机自动化的重要推动力。最近，基于连体网络的跟踪器在离线基准测试中取得了出色的表现。基于 Siamese 网络的跟踪器通常使用经典的深度和广度网络，例如 AlexNet、VggNet 和 ResNet，来提取模板帧和检测帧的特征。然而，由于嵌入式设备的计算能力较差，这些未经修改的模型计算量太大，无法部署在无人机上。在本文中，我们提出了一个设计纤薄主干的指导方针：每一层的输出维度应该小于输入的维度。在该指南的指导下，我们将 AlexNet 的计算要求降低了 59.4%，而跟踪器保持了相当的精度。此外，我们采用无锚网络作为跟踪头，这比基于锚的方法需要更少的计算。基于这些方法，我们的跟踪器在 UAV123 数据集上实现了 60.9% 的 AUC，在 NVIDIA Jetson TX2 上达到了每秒 30 帧，因此可以嵌入到无人机中。据我们所知，它是第一个部署在无人机嵌入式系统上的实时连体跟踪器。该代码可在 GitHub 上获得。