BAROC: Concealing Packet Losses in LSNs with Bimodal Behavior Awareness for Livecast Ingestion

Abstract

The advent of Low-Earth Orbit satellite networks (LSNs), exemplified by initiatives like Starlink, OneWeb and Kuiper, has ushered in a new era of “Internet from Space” global connectivity. Recent studies have shown that LSNs are capable of providing unprecedented download capacity and low latency to support Livecast viewing. However, Livecast ingestion still faces significant challenges, such as limited uplink capacity, bandwidth degradation, and the burst of packet loss due to frequent satellite reallocations, which cause previous recovery and adaptive solutions to be inferior under this new scenario. In this paper, we conduct an in-depth measurement study dedicated to understanding the implications of satellite reallocations, which reveals that the network status during reallocations with network anomalies exhibits a different distribution, leading to bimodal behaviors on the overall network performance. Motivated by this finding, we propose BAROC, a framework that can effectively conceal burst packet losses by combining a novel proposed MTP-Informer with bimodal behavior awareness during satellite reallocation. BAROC enhances video QoE on the server side by addressing the above challenges and jointly determining the optimal video encoding and recovery parameters. Our extensive evaluation shows that BAROC outperforms other video delivery recovery approaches, achieving an average PSNR improvement of 1.95 dB and a maximum of 3.44 dB, along with enhancements in frame rate and parity packet utilization. Additionally, a comprehensive ablation study is conducted to assess the effectiveness of MTP-Informer and components in BAROC.

以Starlink、OneWeb和Kuiper等项目为代表的低地球轨道卫星网络（LSN）的出现，开启了“太空互联网”全球互联的新纪元。近期研究表明，LEO卫星网络能够提供前所未有的下载容量和低延迟，足以支持直播观看。然而，直播的采集（ingestion）过程仍面临巨大挑战，如上行链路容量受限、带宽劣化，以及因频繁的卫星切换（satellite reallocations）导致的突发性丢包。这些挑战使得以往的恢复与自适应解决方案在这一新场景下性能不佳。

在本文中，我们进行了一项深入的测量研究，旨在理解卫星切换所带来的影响。该研究揭示，在发生网络异常的切换期间，网络状态呈现出一种不同的分布，从而导致整体网络性能表现出双模态行为（bimodal behaviors）。受此发现启发，我们提出了BAROC，一个能够有效掩盖突发性丢包的框架。该框架在卫星切换期间，将我们新提出的MTP-Informer模块与对双模态行为的感知能力相结合。BAROC通过应对上述挑战，并协同确定最优的视频编码与恢复参数，从而在服务器端提升视频体验质量（QoE）。

我们广泛的评估结果表明，BAROC的性能优于其他视频传输恢复方法，峰值信噪比（PSNR）平均提升了1.95 dB，最高可达3.44 dB，同时在帧率和冗余包利用率方面也有所增强。此外，我们还进行了一项全面的消融研究（ablation study），以评估BAROC框架中MTP-Informer及其他组件的有效性。

Introduction

Since Starlink first started the commercial service of LowEarth Orbit (LEO) satellite constellation, both industry and academia have witnessed its impressive network performance and noticed its great potential of achieving global Internet coverage in the upcoming era of 6G and beyond. After establishing saturated coverage in most urban areas, the leader operator, Starlink, OneWeb, and Kuiper, expanded its service to include land mobility, maritime, and aviation to realize the proposed Space–Air–Ground network. One of the primary applications benefiting from the expansion of service areas is Livecast, where streamers are typically limited to urban areas because of the need for stable, high-capacity networks. With the expansion of LEO Satellite Networks (LSNs), anytime and anywhere real-time video task applications, such as outdoor Livecasts, wilderness rescues, and surveillance, are becoming increasingly feasible.

In a typical Livecast scenario, streamers first capture highquality video content, upload it to the cloud, and then distribute it to viewers through Content Delivery Networks. Despite the high downlink capacity of LSNs, which can reach around 200Mbps and latency of approximately 50ms, capable of real-time, high-quality video viewing [1], [2], [3], [4], [5], Livecast ingestion still encounter significant challenges due to substantial performance disparities between LSN’s uplink and downlink, as well as higher delivery requirements. Firstly, since each LEO satellite can only cover a small area and moves quickly, it is needed to switch and relay the signal to the next satellite to maintain a “seamless” connection with the user equipment (UE) (as shown in Fig. 1). Frequent satellite reallocations lead to bursts of packet loss and sudden bandwidth degradation [1], [6], [7], increasing network unpredictability and degrading Livecast ingestion performance. While initial efforts have been made to mitigate such impacts for the downlink of Livecast [7], [8], [9], [10], [11], the more challenging issues of real-time ingestion caused by satellite reallocation remains largely unexplored.

Secondly, due to the limited antenna size and power supply employed in the UE, the UE has less transmitting signal power compared to the satellites. For instance, the uplink capacity is 15x smaller than the downlink capacity, which is only around 15Mbps. Thirdly, from the application perspective, the encoding buffer on the server side is much smaller than the video content buffer on the viewer’s side, meaning any packet loss in the uplink will quickly halt the decoding and distribution pipeline on the server, affecting all downstream viewers. Therefore, strategies like slowing down playback speed to extend buffer duration [10], [11] are not feasible for the streamer as the content creator.

In this work, we conduct an in-depth measurement study dedicated to understanding the implications of satellite reallocations, which reveals that the network status during reallocations with network anomalies exhibits a different distribution, leading to bimodal behaviors on the overall network performance. Motivated by this finding, we propose a bimodal behavior aware loss concealing (BAROC) framework for Livecast ingestion. To address packet loss during Livecast, the most common methods are packet- or frame-level recovery [12], [13], [14], [15], as re-transmission is generally impractical due to Livecast latency constraints. These methods usually rely on an accurate predictor to determine the appropriate ratio of redundant (a.k.a parity) packets or require sufficient computation power and time for high-quality frame recovery using Neural encoders. Due to the aforementioned challenges inherently from LSNs, these prerequisites are difficult to meet. To meet the constrained time requirement from the application perspective, an efficient recovery method, such as the forward error correction (FEC) mechanism is recommended. Yet, considering the inclusion of extra parity packets for recovering, and the limited uplink capacity mentioned in LSNs, an accurate LSN-specific network predictor and a more efficient video quality scheduler are also required to maintain both high video quality and packet recovery ratio.

To this end, in the BAROC framework, we propose an enhanced Transformer predictor [16], [17], Multi-Task Probabilistic Informer (MTP-Informer), and with a video quality scheduler that jointly resolves the distribution convolution problem introduced by bimodal behavior and bitrate variance. This framework maximizes the video quality on the server side by jointly adjusting the FEC ratio (the ratio of parity packets) and video encoding parameters on the streamer side considering both satellite reallocation, limited uplink capacity, and variance of video encoded bitrate.

To adapt to the bimodal behavior of the LSNs, the MTPInformer models the probability distribution of the bimodal network metrics, rather than relying solely on single-point observations. It predicts the future probability distribution, providing a more informative prediction for later video scheduling. This approach makes the BAROC aware of less frequent packet loss bursts and also more resilient to inaccurate forecasts. We then formulate a convolution calculation problem that combines the probability distribution of bimodal behaviors and the distribution arising from bitrate variance and resolves with the proposed video quality scheduler.

To evaluate the performance of BAROC, we compare the proposed solution with other state-of-the-art video delivery recovery methods using real-world LSN traces with multiple videos incorporating typical Livecast scenarios. Our emulated experiment shows that BAROC demonstrates improvements in terms of Peak signal-to-noise ratio (PSNR), frame rate, parity packet utilization ratio, and recovery ratio. In detail, BAROC achieves an average PSNR improvement of 1.95 dB, with a maximum of 3.44 dB, and an average increase of 13.54% in parity packet utilization. Additionally, we conduct an ablation study to solely evaluate the effectiveness of components within the BAROC framework. The contributions of this paper can be summarized as follows.

• We conduct a measurement study focused on the impact of satellite reallocations, uncovering the underlying bimodal behaviors. The fluctuating packet loss and bandwidth during these reallocations highlight the need for a novel Livecast ingestion framework tailored to LSNs. (§II)

• We propose the enhanced MTP-Informer specifically for LSNs, which predicts the probability distribution of network metrics based on historical satellite reallocation effect and bimodal behavior. (§III)

• We formulate the distribution convolution problem and introduce a video quality scheduler that jointly determines the optimal video quality and FEC parameters based on the predicted bandwidth and packet loss distributions. (§IV)

• We evaluate the performance of BAROC with other state-of-the-art real-time video delivery recovery methods (R-FEC[13], LightFEC [14], and FBRA [18]) and demonstrate its superiority in terms of multiple evaluation metrics. (§V)

自Starlink首次启动其低地球轨道（LEO）卫星星座的商业服务以来，工业界和学术界共同见证了其卓越的网络性能，并注意到了其在即将到来的6G及未来时代实现全球互联网覆盖的巨大潜力。在大多数城市地区建立起饱和覆盖后，以Starlink、OneWeb和Kuiper为首的运营商已将其服务扩展至陆地移动、海事和航空领域，以实现所提出的“天空地一体化网络”（Space–Air–Ground network）。从服务区域扩张中受益的主要应用之一是直播（Livecast），因为该应用通常需要稳定、高容量的网络，从而使得主播大多局限于城市地区。随着LEO卫星网络（LSNs）的扩张，随时随地的实时视频任务应用，如户外直播、野外救援和监控等，正变得日益可行。

在典型的直播场景中，主播首先采集高质量的视频内容，将其上传至云端，然后通过内容分发网络（CDN）分发给观众。尽管LEO卫星网络具有很高的下行链路容量（可达约200Mbps）和较低的延迟（约50ms），足以支持实时高质量的视频观看[1], [2], [3], [4], [5]，但直播采集（Livecast ingestion）过程仍面临着巨大挑战。这源于其上下行链路之间显著的性能差异，以及更高的传输要求。首先，由于每颗LEO卫星仅能覆盖一小片区域且移动迅速，信号需要在下一颗卫星间进行切换和中继，以维持与用户设备（UE）的“无缝”连接（如图1所示）。频繁的卫星切换（satellite reallocations）会导致突发性丢包和带宽骤降[1], [6], [7]，这增加了网络的不可预测性，并降低了直播采集的性能。尽管已有初步工作尝试缓解此类问题对直播下行链路的影响[7], [8], [9], [10], [11]，但由卫星切换所引发的、更具挑战性的实时采集问题在很大程度上仍未被探索。

其次，由于用户设备（UE）所采用的天线尺寸和电源供应有限，其信号发射功率远小于卫星。例如，上行链路容量比下行链路小15倍，仅为约15Mbps。第三，从应用角度看，服务器端的编码缓冲区远小于观众端的视频内容缓冲区，这意味着上行链路中的任何丢包都将迅速中断服务器端的解码与分发流水线，从而影响所有下游观众。因此，诸如减慢播放速度以延长缓冲时长的策略[10], [11]，对于作为内容创作者的主播而言是不可行的。

在本文中，我们进行了一项深入的测量研究，旨在理解卫星切换所带来的影响。该研究揭示，在发生网络异常的切换期间，网络状态呈现出一种不同的分布，从而导致整体网络性能表现出双模态行为（bimodal behaviors）。受此发现启发，我们提出了一个具备双模态行为感知的丢包隐藏框架（BAROC）用于直播采集。为解决直播中的丢包问题，最常用的方法是包级或帧级的恢复[12], [13], [14], [15]，因为重传通常因直播的延迟约束而不可行。这些方法通常依赖一个精确的预测器来确定合适的冗余（即校验）包比例，或需要足够的计算能力和时间来使用神经编码器进行高质量的帧恢复。由于LEO卫星网络固有的上述挑战，这些先决条件难以满足。为满足应用层面的严格时间要求，推荐使用如前向纠错（FEC）机制等高效的恢复方法。然而，考虑到为恢复数据需额外加入冗余包，以及LEO卫星网络有限的上行容量，一个精确的、为LEO网络专用的网络预测器和一个更高效的视频质量调度器也同样是必需的，以同时维持高视频质量和高丢包恢复率。

为此，在BAROC框架中，我们提出了一个增强版的Transformer预测器[16], [17]，即多任务概率Informer（MTP-Informer），并配合一个视频质量调度器。该调度器能够协同解决由双模态行为和码率变化所引入的分布卷积问题。该框架通过在主播端协同调整FEC冗余度（冗余包的比例）和视频编码参数，同时考虑卫星切换、有限的上行容量以及视频编码码率的变化，从而最大化服务器端的视频质量。

为适应LEO卫星网络的双模态行为，MTP-Informer对双模态网络指标的概率分布进行建模，而非仅仅依赖单点观测值。它预测未来的概率分布，为后续的视频调度提供信息更丰富的预测。这种方法使BAROC能够感知到频率较低的突发性丢包，并且对不准确的预测也更具鲁棒性。随后，我们构建了一个卷积计算问题，该问题结合了双模态行为的概率分布和由码率变化产生的分布，并利用所提出的视频质量调度器进行求解。

为评估BAROC的性能，我们利用真实的LEO卫星网络轨迹和包含典型直播场景的多个视频，将所提出的解决方案与其他先进的视频传输恢复方法进行了比较。我们的仿真实验表明，BAROC在峰值信噪比（PSNR）、帧率、冗余包利用率和恢复率方面均表现出提升。具体而言，BAROC的PSNR平均提升了1.95 dB，最高可达3.44 dB，冗余包利用率平均增加了13.54%。此外，我们还进行了一项消融研究（ablation study），以单独评估BAROC框架内各组件的有效性。本文的贡献可总结如下：

• 我们进行了一项专注于卫星切换影响的测量研究，揭示了其潜在的双模态行为。在切换期间，波动的丢包率和带宽凸显了设计一个针对LEO卫星网络、新颖的直播采集框架的必要性。（§II）
• 我们提出了专为LEO卫星网络设计的增强型MTP-Informer模块，它能够基于历史上的卫星切换影响和双模态行为，来预测网络指标的概率分布。（§III）
• 我们构建了 “分布卷积”问题模型，并引入了一个视频质量调度器。该调度器能够基于预测的带宽和丢包率分布，协同确定最优的视频质量与前向纠错（FEC）参数。（§IV）
• 我们将BAROC的性能与其他先进的实时视频传输恢复方法（R-FEC [13], LightFEC [14], and FBRA [18]）进行了比较，并证明了其在多项评估指标上的优越性。（§V）

Background and Motivation

LEO networks and satellite reallocations. For most LSN operators, such as Starlink and OneWeb, their commercial UE is outfitted with a single-phased array antenna for sending and receiving signals. During UE-Satellite (UE-Sat) link rescheduling, the UE requests a new spot beam from the incoming satellite and releases the current one upon confirmation. During this beam switch, packets in the UE-Sat link, such as packet 3 in Fig. 1, may be lost due to delays in routing table updates or resource reconfiguration [1], [7], [19].

In practice, we observe that bursts of lost packets can affect tens or even hundreds of packets sequentially, which will also significantly degrade the bandwidth. Both documentation and measurement work [1], [6], [20] have confirmed that the Starlink employs a global controller for UE-Sat link reallocation schedule, where the UE-Sat link will be rescheduled every 15 second, and specifically at 12th, 27th, 42th, and 57th second in each minute for all UEs. While the reallocation scheduling is known, the actual impact of reallocation on end-to-end network performance remains ambiguous. The reallocation effect can exhibit either significant network degradation or minimal visible impact most of the time, resulting in bimodal behaviors, as demonstrated by our measurement study to be discussed next. Furthermore, compared to reallocation in cellular or WiFi networks, satellite reallocation has more pronounced consequences. For example, in LSNs, the satellite is mobile while the user (UE) remains relatively static, causing all users to experience periodic reallocation effects during usage. For cellular or WiFi networks, reallocations only affect users who are moving away from the current access point.

LEO 网络与卫星重分配问题。 对于大多数低轨卫星网络（LSN）运营商（如 Starlink 与 OneWeb）而言，其商用用户终端（UE）通常配备单个相控阵天线用于信号的收发。在 UE 与卫星（UE-Sat）链路重分配过程中，UE 会向即将接入的卫星请求新的波束资源，并在确认后释放当前波束。在这一波束切换阶段，UE-Sat 链路中传输的分组（如图 1 中的分组 3）可能因路由表更新或资源重新配置所引发的延迟而丢失 [1], [7], [19]。

实践中我们观察到，这类分组丢失通常呈现突发性，可连续影响数十甚至上百个分组，从而导致带宽显著下降。已有文档资料与测量研究 [1], [6], [20] 表明，Starlink 采用全局控制器以统一调度 UE-Sat 链路的重分配，并且该重分配过程在每分钟的第 12 秒、27 秒、42 秒与 57 秒统一触发，周期为 15 秒。

然而， 尽管重分配的调度时刻是已知的，其对端到端网络性能的实际影响仍缺乏明确认知 。根据我们后续测量研究的结果，链路重分配的影响存在显著的双峰性（bimodal behavior）：即多数情况下几乎无感知影响，而在某些重分配阶段却会显著恶化网络性能。此外，与蜂窝或 WiFi 网络中的重分配相比，卫星网络中的重分配后果更为严重。 在 LSN 中，由于卫星持续移动而用户终端相对静止，所有用户都会周期性地经历链路重分配带来的影响；而蜂窝或 WiFi 网络中，链路重分配通常只影响正在移动的用户。

Measurements on bimodal behaviors in LSNs. Prior works [1], [2], [3], [4] have presented a comprehensive network evaluation on LSNs, our focus is specifically on examining LSNs’ performance degradation during reallocation, particularly concerning bandwidth, latency, and packet loss ratio. For a typical Livecast scenario, the recommended maximum packet loss ratio is less than 2% according to Zoom’s User guideline 1 . Thus, we define a period of 1-second as network anomaly if we observe the average packet loss ratio is larger than 2%.

During our week-long measurement (detailed measurement setup in §V-A), 30.73% of reallocation periods were marked as network anomaly, compared to only 4.32% normal period marked as network anomaly. Fig. 2 illustrates the average performance degradation during periods of reallocation with network anomalies compared to normal periods. The latency and packet loss ratio both show significant increases, rising by factors of 4.49 and 16, respectively, while the bandwidth also experiences a decrease of 24%. The results above indicate that not all satellite reallocations affect the application level. However, when they do, the impact can significantly degrade the performance of Livecast applications. Moreover, Fig. 3 illustrates the overall probability distribution during the normal period and the reallocation period with network anomalies. The figure highlights significant differences in both the mean and distribution of bandwidth and packet loss ratios between these two periods. This interleaved discrepancy, characterized by a minimum switch time of 15 seconds, causes abrupt fluctuations in the LSNs, rendering general predictors ineffective, as shown in our subsequent evaluation.

为了决策最优的视频质量与前向纠错（FEC）冗余比率，我们结合历史网络性能数据与链路重分配的领域知识，对未来重分配事件对网络的潜在影响进行预测。该方法在无需深入分析复杂卫星拓扑的同时，仍能有效捕捉重分配的关键特征。进一步地，我们根据重分配期间是否伴随网络异常，对数据进行分组并将该标签作为新的输入特征引入模型中，从而增强模型识别时间序列中双峰特性的能力。这一改进使得模型能够输出包含更多信息量的分布预测结果，而非单一的点值预测。

LSN 中双峰行为的测量分析。 虽然已有工作 [1]–[4] 对 LSN 网络进行了全面性能评估， 本文聚焦于链路重分配期间 LSN 的性能退化，重点分析带宽、时延与分组丢失率。 在典型的实时视频场景中，根据 Zoom 的用户指南，推荐的最大分组丢失率不应超过 2%。因此，我们将 1 秒为单位的时间段定义为“网络异常”期，若该时段的平均丢包率超过 2%。

在为期一周的测量过程中（具体设置详见 §V-A），我们发现有 30.73% 的重分配期被判定为网络异常，而正常期仅有 4.32%。图 2 展示了网络异常重分配期与正常期之间的平均性能差异。其中时延与丢包率分别上升了 4.49 倍与 16 倍，带宽下降了 24%。这一结果表明，尽管并非所有卫星重分配都会影响应用层性能，但一旦发生，其对直播类应用的影响是极其显著的。进一步地，图 3 展示了正常期与网络异常重分配期的整体概率分布，突显了两者在带宽与丢包率分布上的显著差异。此种交替出现的突变行为（最小切换间隔为 15 秒）导致 LSN 中网络状态剧烈波动，从而使得传统的点值预测方法难以胜任，如我们后续评估中所示。

Motivated by the above measurement and analysis, we advocate:

(i) Probabilistic prediction for LSNs. Considering the challenges outlined above and the need to understand the pattern of UE-Sat reallocation to decide the suitable video quality and FEC ratio, we forecast the impact of future reallocation relying on historical network performance coupled with domain knowledge of reallocation characteristics. This method can bypass the analysis of complex and hidden satellite topology while remaining attuned to critical reallocation details. Additionally, we classify our data into distinct groups according to their association with network anomalies during reallocation and incorporate this information as a new input to the model. This modification enables the model to identify the presence of bimodal behavior in the time-series prediction task and generate more informative distribution predictions rather than just single-point predictions.

（i）LSN 的概率性预测

鉴于 UE-Sat 重分配模式对视频传输的影响巨大，且其规律可被预测，我们提出一种结合历史网络性能数据与重分配机制知识的预测方法，以推断未来重分配可能带来的影响。该方法避免了复杂的卫星轨道与链路拓扑建模，聚焦于对影响行为本身的建模。进一步地，我们引入重分配是否伴随网络异常作为分类特征输入模型，使其能够识别时间序列预测中存在的双峰性，并输出分布式预测结果，提升对重分配行为的可解释性与鲁棒性。 s (ii) Scheduler for distribution convolution. Due to the high packet loss ratio during the reallocation period, adding parity packets further limits the already constrained uplink capacity, severely restricting the available goodput for video data in LSN’s uplink. To address this, Variable Bitrate (VBR) encoding is necessary to achieve higher compression ratios compared to the general Constant Bitrate (CBR) encoding used in Livecast ingestion. As shown in our video set in §V, VBR encoding provides a PSNR improvement of 2.35 ± 3.27 dB over CBR. In VBR, video quality is adjusted using the Constant Rate Factor (CRF), but the actual bitrate depends on texture complexity and motion speed. The variability in chunk sizes caused by VBR poses an additional challenge in determining the optimal FEC ratio, as it is closely tied to each video’s chunk size. To address this, we introduce a CRF-bitrate distribution construction algorithm that analyzes real-time bitrate probability distributions and maps them to CRF values. Additionally, the video quality scheduler optimizes video parameters by jointly considering probabilistic predictions and bitrate distributions from video encoding.

（ii）用于分布卷积的调度器设计

由于链路重分配期间的高丢包率，传统通过增加冗余包的前向纠错方法进一步压缩了已受限的上行链路容量，显著降低了可用于视频传输的有效带宽（goodput）。为缓解此问题，我们采用可变码率（VBR）编码替代传统的恒定码率（CBR）编码，以实现更高的压缩率。我们的视频测试集（详见 §V）表明，VBR 编码相比 CBR 平均可提升 2.35 ± 3.27 dB 的 PSNR。在 VBR 编码中，视频质量由恒定速率因子（CRF）控制，而实际码率则受纹理复杂度与运动速度影响。由此带来的码率波动使得 FEC 比率的最优选择更加复杂。对此，我们设计了一种 CRF-码率分布 构建算法，能够分析实时编码后的码率分布，并映射至合适的 CRF 值。视频调度器则基于概率预测结果与视频编码生成的码率分布，联合优化视频参数，从而实现端到端的自适应传输优化。

TL; DR

聚焦于在 User-Sat 切换时, 优化上行链路传输

范围比较窄, 考察的是 "重分配" 时刻引起的 "User-Sat切换", 这个 reallocation 是可预测的。原因: 运营商使用一个全局控制器来安排用户设备（UE）与卫星之间的链路重分配, 个重分配调度是周期性的，用户与卫星的链路会每15秒重新调度一次

1. 狼人: 周期性的 reallocation 会引起周期性地 “抖一抖”
   1. 正常模式 - 异常模式: 这种“要么没事，要么事很大”的两极化表现，就是所谓的“双模态行为”
   2. 这种切换情况会引发两种可能的结果:
      1. 网络非常宽裕与弹性: 平稳过渡，几乎没有影响
      2. 网络资源本身就比较紧张: 剧烈抖动, 丢包!
2. 预言家: 一个专为卫星网络设计的、名为 MTP-Informer 的新型网络预测器
   1. 信令形如: "xdm, 未来的第15s会发生一次reallocation切换, 30%的概率会发生剧烈抖动，届时带宽可能会掉到5Mbps以下. 丢包率会飙升到10%; 另外70%的概率是平稳过渡，网络几乎不受影响"
3. 骑士: 视频质量调度器
   1. 给出一个 "最优打包运输方案"
      1. 稍微降低一点画质
      2. 增加前向纠错FEC的冗余包 (“双保险”)

注意:

1. 优化的是上行链路:
   • 特点: 信号发射功率远小于卫星 (上行链路容量比下行链路小15倍)
2. 概率预测:
   • MTP - Transformers ("Attention is All You Need")
3. 自适应码率:
   • 传统通过增加冗余包的前向纠错方法进一步压缩了已受限的上行链路容量
   • 我们采用可变码率: VBR

补充知识: WebRTC相关过程

(1) 码率: Bitrate / BR

1. 单位: bps
2. 含义: BR越高, 每秒钟用来描述视频画面的数据就越多

(2) CRF: 质量目标

1. 使用模式: 不是直接告诉编码器“给我 5 Mbps 的码率”, 而是告诉它 “我想要 CRF=23 这个级别的画质”
2. CRF值越低, 画质越高, 码率也越高

主播 side

(3) 分辨率 (Resolution): 一个静态视频切片中, 像素点的数目

分辨率越高，画面越清晰，但原始数据量也越大

(4) 帧率 (Frame Rate, FPS): 每秒钟显示的图像帧数

1. 例如 60 FPS 就是一秒钟由60张连续的静态图片组成
2. 帧率越高，视频看起来就越流畅

(5) 编解码器:

1. Encoder: 负责将采集到的巨大原始视频数据进行压缩，去掉冗余信息，把它变成适合网络传输的码流
2. Decoder: 在接收端，负责将接收到的码流解压缩，还原成我们可以观看的视频画面

(6) 画面类别 (GOP, Group of Pictures):

1. I 帧: 一张完整的、自给自足的图像。解码时不需要参考其他任何帧
2. P 帧: 只记录与前一帧相比发生变化的差异部分。例如，主播眨了一下眼，P帧就只记录眼睛部分的变化，背景不变
3. B 帧: 双向预测 (暂时先不管)

网络传输 viewpoint

(7) 抖动 (Jitter): 指数据包到达时间的不规律性

1. 比如第一个包花了20ms，第二个包花了50ms，第三个包花了30ms。这种时快时慢的现象就是抖动
2. 接收端需要一个抖动缓冲 (Jitter Buffer) 来把这些乱序、乱速到达的包重新排好队，再平滑地播放出去

(8) 错误恢复机制 (Error Recovery):

1. ARQ机制 (Automatic Repeat Request): 自动重传请求, "我没收到第5个包，请你重发一次"
2. FEC机制 (Forward Error Correction):
   • 我给你发10个数据包的同时，额外再发2个冗余包（Parity Packet）
   • 这两个冗余包是根据前10个包计算出来的。路上就算丢了任意2个包，你也能用剩下的包和冗-余包把丢失的包算出来，无需重传