Design Overview

EagleEye is a new constellation organization and operating model that leverages orbital edge computing and constellation design to provide high-coverage, high-resolution data. The viability of EagleEye hinges on the recent maturation of orbital edge computing, effective crosslinking between LEO nanosatellites, and robotics advances that support agile pointing. The key ideas in EagleEye are:

(i) a mixed-resolution leader-follower constellation organization

(ii) actuation-aware scheduling for follower pointing

(iii) target clustering to increase coverage

Mixed-Resolution Leader-Follower Constellations

EagleEye leverages a heterogeneous, mixed-resolution, leaderfollower constellation organization. Leader and follower satellites contain different compute and sensing hardware. Fig. 1a illustrates the EagleEye constellation organization. A leader satellite has a low-resolution imager and compute hardware that enables processing low-resolution frames with high performance. A follower satellite has a high-resolution imager and may or may not include high-performance, computational hardware. Fig. 5 shows how EagleEye constellations differ from existing work, allocating satellites in a constellation into heterogeneous leader-follower groups instead of tasking them homogeneously. Leaders and followers have radio equipment for cross-link [65, 66] and downlink communication. All satellites have ADACS and GPS/GNSS for precise pointing and attitude/orbit determination.

EagleEye 系统采用了一种异构 (heterogeneous)、混合分辨率 (mixed-resolution) 的主从式 (leader-follower) 星座架构。主导卫星 (Leader) 与跟随卫星 (Follower) 搭载了不同的计算硬件和传感硬件。

图1a展示了EagleEye的星座架构。主导卫星配备了低分辨率成像仪和高性能计算硬件，能够高效处理低分辨率图像帧。跟随卫星则配备高分辨率成像仪，并可能（或可能不）包含高性能计算硬件。

图5展示了EagleEye星座与现有工作的不同之处：它将星座内的卫星划分为异构的主从小组，而非对它们进行同构的任务分配。主导卫星和跟随卫星均配备了用于星间链路 (cross-link) 和下行链路 (downlink) 通信的无线电设备。所有卫星都搭载了姿态确定与控制系统 (ADACS) 和全球定位/导航卫星系统 (GPS/GNSS)，以实现精确定向以及姿态和轨道的确定。

EagleEye defines an operating model for a leader-follower constellation. Operationally, a leader satellite images its entire ground track, geo-registering each image with GPS coordinates. The leader processes each image using a pretrained ML model that identifies the targets in the image. A constellation has one or more followers that collectively capture high-resolution images of all targets identified by the leader. The leader distributes the target imaging tasks to the followers. The leader first uses a crosslink to query the position and attitude of each follower. The leader then computes an actuation-aware schedule of image captures. The schedule is a series of pointing and capture actions that a follower should perform to image targets at high-resolution. The actuationaware schedule involves all followers and, when feasible, covers all targets. The leader distributes to each follower its schedule of pointing and capture actions and each follower executes the schedule, capturing and storing the data. Eventually, follower satellites may perform additional onboard processing of the frames or may transmit the captured, high-resolution frames to Earth for consumption by a downstream application. The primary benefit of the EagleEye operating model is higher coverage at high resolution, because the high-resolution followers focus their sensing on areas with targets identified by the leader using low-resolution, highcoverage data.

EagleEye 为这种主从式星座定义了一种运行模型。

在操作上，主导卫星对其整个星下点轨迹 (ground track) 进行成像，并利用GPS坐标对每幅图像进行地理配准 (geo-registering)。

随后，主导卫星使用预训练的机器学习 (ML) 模型处理每幅图像，以识别其中的目标。

一个星座拥有一颗或多颗跟随卫星，它们协同工作，共同捕获由主导卫星识别出的所有目标的高分辨率图像。主导卫星将目标成像任务分配给这些跟随卫星。

首先，主导卫星通过星间链路查询每颗跟随卫星的位置和姿态。接着，主导卫星会计算一个感知驱动能力 (actuation-aware) 的图像捕获调度方案。该调度方案是跟随卫星为拍摄目标高分辨率图像而应执行的一系列指向与捕获动作。这个感知驱动能力的调度方案涉及所有跟随卫星，并在可行的情况下覆盖所有目标。主导卫星向每颗跟随卫星分发其专属的指向与捕获动作调度，每颗跟随卫星则执行该调度，进行图像捕获和数据存储。最终，跟随卫星可以对图像帧进行额外的星上处理，或将捕获的高分辨率图像帧传输回地球，供下游应用 (downstream application) 使用。EagleEye运行模型的主要优势在于，通过让高分辨率的跟随卫星将其探测能力集中在由主导卫星利用低分辨率、高覆盖率数据所识别出的目标区域，从而实现了更高的高分辨率覆盖率。

Challenges of Target Detection & Sensor Scheduling

The goal of the constellation is to efficiently identify targets without exceeding time or energy limits on the leader, and at the same time maximize the number of targets captured by followers. At a high level, the scheduling algorithm achieves this goal by assigning a priority score to each target, based on the confidence with which it was detected, as reported in the output of the target identification ML model. The scheduling’s optimization function is to maximize the sum of priority scores of targets captured by followers in a schedule. As for target detection by the leader, the targets must be large enough for the leader’s low-resolution camera to observe.

星座的目标是在不超过主导卫星时间和能量限制的前提下高效识别目标，同时最大化跟随卫星捕获的目标数量。从宏观层面讲，调度算法通过为每个目标赋予一个优先级分数来实现此目标，该分数基于目标识别ML模型输出的探测置信度。调度的优化函数是最大化调度方案中由跟随卫星捕获的目标的优先级分数之和。对于主导卫星的目标探测而言，目标尺寸必须足够大，以便其低分辨率相机能够观测到。

Challenge 1: Actuation-aware scheduling. The pointing and imaging schedule that the leader produces for each follower must take into account the follower’s position, attitude, and actuation constraints. For each target assigned to a follower, there is a window of time during which that target is available for imaging. The window is defined by the maximum “off-nadir” pointing angle, as shown in Fig. 6 (Left). Once the satellite’s pointing angle exceeds this maximum threshold, the captured images become excessively distorted, rendering them unusable. Also, the schedule should consider the pointing actuation time as shown in Fig. 6 (Right), which limits the number of targets a follower can capture.

主导卫星为每颗跟随卫星生成的指向与成像调度，必须考虑到跟随卫星的位置、姿态及其驱动约束 (actuation constraints)。对于分配给跟随卫星的每个目标，都存在一个可供成像的时间窗口。如图6（左）所示，该窗口由最大 “偏离天底角” (off-nadir) 指向角定义。 一旦卫星的指向角超过此最大阈值，捕获的图像会产生严重畸变 (distorted)，从而变得无法使用。此外，调度还需考虑图6（右）所示的指向驱动时间，这一因素限制了单颗跟随卫星能够捕获的目标数量。

Challenge 2: Limited target detection and scheduling time. Time and energy are the primary limiting factors in a computational nanosatellite [30]. For full ground track coverage, the leader must capture each completely new frame that it observes, which implies a capture cadence of e.g., 15s at 500km with a 100km swath. The leader has 15s to complete target object detection and scheduling. Amplifying the challenge, cubesats usually have limited computing power in embedded CPUs and GPUs. Besides, cubesats get limited energy through solar panel and a single solar panel could only support the satellite computer to run for a portion of its orbit.

时间和能量是计算型纳卫星 (nanosatellite) [30] 的主要限制因素。为了实现对星下点轨迹的完全覆盖，主导卫星必须捕获其观测到的每一个全新视场帧，这意味着在500公里轨道高度和100公里刈幅 (swath) 的情况下，捕获周期约为15秒。 主导卫星必须在这15秒内完成目标物体的探测和调度。 更具挑战性的是，立方星 (cubesat) 的 嵌入式CPU和GPU通常计算能力有限 。此外，立方星通过太阳能电池板获取的能量有限，单个太阳能帆板仅能支持星上计算机在轨道的一部分时段内运行。

While high-accuracy object detection within these time and energy constraints is challenging, recent work [28] provides software solutions to reduce ML execution time in an accuracy-aware manner. These advances make it possible for EagleEye to support actuation scheduling based on ML detection results.

尽管在这些时间和能量约束下实现高精度目标探测极具挑战性，但近期的研究工作 [28] 提供了一些软件解决方案，能够以一种感知精度 (accuracy-aware) 的方式减少ML模型的执行时间。这些技术进步使得EagleEye能够支持基于ML探测结果的驱动调度。

Problem Formulation

TL; DR