Related Work

Our work builds on results in terrestrial edge computing, video processing systems, and orbital edge computing.

Terrestrial edge computing pipelines: Edge computing has been a widely studied topic in terrestrial networks [8,13,17,29, 54,55] for diverse applications such as traffic camera analytics [8,12], augmented reality systems [56], and robotics [63]. These systems tend to push computation as close to the video sensor as possible while being cognizant of the resource constraints of the edge devices. Serval naturally builds on this line of research. Satellites are similar to resource-constrained mobile devices with relatively weak connectivity to the cloud. The LEO setting presents the additional challenge that the sensor devices (satellites) themselves are moving. Serval addresses this by leveraging the predictable orbital paths of satellites, query filters and bifurcation, and using auxiliary information.

地面边缘计算流水线：

地面网络中的边缘计算已是一个被广泛研究的课题 [8, 13, 17, 29, 54, 55]，其应用涵盖交通摄像头分析 [8, 12]、增强现实系统 [56] 和机器人技术 [63] 等多个领域。 这些系统倾向于在认知到边缘设备资源受限的同时，将计算尽可能地推向靠近视频传感器的位置。 Serval自然地建立在这一研究方向之上。卫星类似于资源受限且与云端连接较弱的移动设备。低地球轨道（LEO）环境带来了额外的挑战，即传感器设备（卫星）自身处于移动之中。Serval通过利用卫星的可预测轨道路径、查询滤波器与处理分流（bifurcation），以及使用辅助信息来应对这一挑战。

Video processing systems: Much recent work [6,8,13,39,48, 65] has focused on improving the execution of video analytics pipelines on edge devices. These systems consider different aspects of optimizing video analytics such as efficient model retraining [13], model merging for efficient execution on edge GPUs [48], etc. This line of work makes varying assumptions about the availability of compute resources such as powerful GPUs and continuous connectivity with the cloud—these luxuries are not available on satellites. Nevertheless, ideas from video processing systems are complementary to Serval, and our work opens an avenue to explore such future directions. For instance, model merging, an idea from video processing systems, can be useful if the models (inside the filters of different Serval queries) share a lot of common layers.

视频处理系统：

近期的许多工作 [6, 8, 13, 39, 48, 65] 专注于提升视频分析流水线在边缘设备上的执行效率。这些系统考虑了优化视频分析的不同方面，例如高效的模型重训练 [13]、为在边缘GPU上高效执行而进行的模型融合 [48] 等。这一系列工作对计算资源的可用性（如强大的GPU）和与云端的持续连接性做出了不同程度的假设——而这些奢侈的条件在卫星上并不存在。尽管如此，来自视频处理系统的思想与Serval是互补的，我们的工作为探索这些未来方向开辟了道路。例如，模型融合这一思想，如果不同Serval查询的滤波器内部共享大量共同的层，那么它将非常有用。

Satellite edge computing: Traditional satellites packaged radiation-hardened specialized hardware [10]. Due to the lower orbits of LEOSats which suffer much less radiation exposure, and for economies of scale in manufacturing, there has been a move towards general-purpose hardware for LEOSats [28]. This has blossomed research in satellite edge computing systems [15,16,22,23,43–45]. Some papers propose new edge-enabled functionalities for communication megaconstellations [15,16], e.g., deploying content delivery networks in space to improve network performance. This work is independent of Serval due to its focus on a different class of satellites. [45] and [44] evaluate the performance of compression techniques and Machine learning models on satellite-compatible hardware such as NVIDIA Jetson Nano and/or NVIDIA Jetson AGX. We believe such optimizations demonstrate the feasibility of deploying Machine Learning workloads on satellites, and such architectural optimizations (including from other domains like approximate computing) can be applied orthogonally to Serval to improve performance.

在轨边缘计算:

传统卫星通常搭载抗辐射的专用硬件 [10]。由于低轨卫星（LEOSats）遭受的辐射暴露要少得多，并出于制造业的规模经济考虑，低轨卫星已转向使用通用硬件 [28]。这催生了在轨边缘计算系统的研究热潮 [15, 16, 22, 23, 43–45]。一些论文为通信巨型星座提出了新的边缘赋能功能 [15, 16]，例如在太空中部署内容分发网络以提升网络性能。这项工作因其关注的是不同类型的卫星，故与Serval的研究相互独立。[45]和[44]评估了压缩技术和机器学习模型在卫星兼容硬件（如NVIDIA Jetson Nano和/或NVIDIA Jetson AGX）上的性能。我们相信，这类优化证明了在卫星上部署机器学习负载的可行性，并且这类架构层面的优化（包括来自近似计算等其他领域的优化）可以正交地（orthogonally）应用于Serval以提升其性能。

Our work is closest to Orbital Edge Computing (OEC [23]) and Kodan [22]. Both OEC and Kodan aim to reduce the amount of satellite imagery transferred to Earth by enabling early rejection of imagery that is not considered useful. While OEC reimagines different satellites in a constellation as a computational nanosatellite pipelines that seamlessly distribute computational tasks, Kodan focuses more on the computation at each satellite. In contrast to both these works, Serval does not discard any images on the satellite, and focuses on reducing the delay of latency-sensitive images via prioritization and reordering. This has the advantage that post facto queries on historical can be performed, e.g., in a recent high-profile incident, Planet used historical satellite imagery to trace the historical motion of a Chinese balloon entering into the US airspace from its origin to destination [5]. Kodan’s contribution is orthorgonal to ours, since Kodan aims to train the optimal neural network model for specific user applications, while we focus on how to optimally schedule compute on satellites and ground stations given a fixed neural network. Hence Kodan can be used alongside Serval.

我们的工作与在轨边缘计算（OEC [23]）和Kodan [22]最为接近。OEC和Kodan的目标都是通过对被认为无用的图像进行早期拒绝，来减少传输到地球的卫星影像数据量。

OEC将星座中的不同卫星重构为一条能够无缝分发计算任务的计算型纳卫星流水线，而Kodan则更侧重于单颗卫星上的计算。

与这两项工作相比，Serval不在卫星上丢弃任何图像，而是通过优先级划分与重排序来专注于降低延迟敏感图像的交付延迟。

这样做的好处是，系统可以对历史数据执行事后查询（post facto queries）。例如，在近期一个备受瞩目的事件中，Planet公司便利用历史卫星影像，追溯了一个中国气球从起点到终点进入美国空域的历史轨迹 [5]。Kodan的贡献与我们的工作是正交的，因为Kodan旨在为特定的用户应用训练最优的神经网络模型，而我们则专注于在给定一个固定的神经网络模型后，如何优化调度卫星和地面站上的计算任务。因此，Koden可以与Serval协同使用。