A Survey on Efficient Edge-Based Video Streaming: Integrating AI-Powered Upscaling, Adaptive Delivery, and Latency Reduction

References

[1] Approach / Contribution

[2] Key Findings

[3] Limitations

[4] Future Directions

[5] [1] Li et al. (2020)

[6] DRL-based ABR adaptation

[7] Reduced buffering, improved QoE

[8] High computation cost, poor generalization

[9] Multi-agent DRL for multi-user

[10] [2] Xu et al. (2018)

[11] HTTP/2 for streaming

[12] Lower connection overhead, prioritized segments

[13] Limited backward compatibility

[14] HTTP/3, QUIC integration

[15] [3] Mengistu (2023)

[16] NVIDIA DLSS AI upscaling

[17] Higher frame rates, stable quality

[18] Proprietary hardware dependence

[19] Open-source, portable upscalers

[20] [4] Lundkvist (2021)

[21] CNN-based upscaling (SVT)

[22] Better quality at lower bitrates

[23] High compute cost, workflow integration issues

[24] Hybrid cloud–edge, optimized CNN

[25] [5] Hu et al. (2023)

[26] Survey of edge analytics

[27] Highlighted federated learning, edge ML

[28] Lack of standards, privacy gaps

[29] Interoperable, secure frameworks

[30] [6] Nassisid et al. (2024)

[31] 6G adaptive framework

[32] Lower latency, predictive buffer control

[33] No real-world validation

[34] Field trials, semantic protocols

[35] [7] Choi et al. (2019)

[36] MDP + Lyapunov optimization

[37] Reduced rebuffering, improved QoE

[38] Requires perfect network knowledge

[39] Handle partial observability

[40] [8] Chen et al. (2024)

[41] SODA ABR (convex optimization)

[42] Stable playback, fewer bitrate shifts

[43] Weak in interference-heavy networks

[44] Network slicing, cross-layer design

[45] [9] Yeregui et al. (2025)

[46] 5G RAN predictive model

[47] Accurate under controlled settings

[48] Fails with mobility/handoffs

[49] Context-aware, mobility-integrated

[50] [10] Dharuman et al. (2023)

[51] Edge + ML for HLS/DASH

[52] Reduced CDN load, improved efficiency

[53] Cache misses in flash crowds

[54] Hybrid AI + rule-based fallback

[55] [11] STL (2025)

[56] Edge for sports streaming

[57] 47% latency reduction, interactivity

[58] High energy/cost overhead

[59] Renewable integration, scaling

[60] [12] Muvi (2025)

[61] Edge micro-DCs, Kubernetes

[62] 200 ms latency, dynamic scaling

[63] Cold-start, surge inefficiency

[64] Predictive auto-scaling

[65] [13] Bilal & Erbad (2017)

[66] Edge transcoding (Docker)

[67] 60% bandwidth saving, multi-tenant

[68] GPU sharing risks

[69] TEEs, isolation, optimization

[70] [14] Choi et al. (2025)

[71] DRL + D2D in vehicular nets

[72] 33% better reliability, fewer interruptions

[73] Weak in adverse weather

[74] Channel coding, env. sensing

[75] [15] Ghosh et al. (2023)

[76] REACT edge–cloud analytics

[77] 18% fewer false positives

[78] Added latency due to encryption

[79] Lightweight privacy techniques

[80] [16] NSF (2025)

[81] Buffer + SVM QoE prediction

[82] 92% user satisfaction

[83] Device heterogeneity issues

[84] Content-aware, device-specific

[85] [17] Bilal et al. (2018)

[86] FoV-aware 360° caching

[87] 55% bandwidth saving

[88] Low accuracy in dynamic VR

[89] Multimodal gaze prediction

[90] [18] ElasticEdge (2022)

[91] Kubernetes elastic edge

[92] 40% latency reduction

[93] Cold-start delays (15 s)

[94] Pre-warming, proactive allocation

[95] Table I - Comparative Summary of Edge-Based Video Streaming Research

[96] III. PROPOSED MODEL

[97] The proposed architecture integrates computational intelligence techniques at the edge layer to optimize video delivery across four modules: AI-based upscaling, adaptive delivery engine, intelligent edge node management, and latency optimization protocols. The framework aims to minimize bandwidth overhead, reduce latency, and improve Quality of Experience (QoE). A conceptual block diagram is shown in Fig. 1.

[98] Fig. 1 - Block Diagram of Proposed Computational-Intelligence–Enabled Edge Streaming Model

[99] A. AI-Based Edge Upscaling

[100] Video frames are transmitted at reduced resolution from the origin server to conserve bandwidth.

[101] At the edge, deep convolutional super-resolution models such as Real-ESRGAN are employed to reconstruct high-resolution frames.

[102] The upscaler leverages motion-compensated input to maintain temporal consistency across frames, preventing flicker or visual artifacts during fast-motion sequences.

[103] This module enables bandwidth savings of up to 30–40% while preserving perceptual quality, based on benchmarks reported in prior studies.

[104] B. Adaptive Delivery Engine with Reinforcement Learning

[105] The adaptive engine continuously monitors network throughput, playback buffer occupancy, and device capability.

[106] A Deep Reinforcement Learning (DRL) agent selects optimal bitrate/resolution pairs using a state–action–reward loop:

[107] State: current bandwidth, buffer level, latency estimate.

[108] Action: select next video segment bitrate/resolution.

[109] Reward: QoE score, combining rebuffering penalty, visual quality, and stability.

[110] Unlike rule-based ABR, the DRL approach adapts proactively to varying conditions, reducing abrupt bitrate shifts and buffering events.

[111] C. Edge Node Caching and Resource Management

[112] Edge nodes handle local caching and transcoding of popular content to reduce backhaul traffic.

[113] A predictive caching algorithm is applied, where machine learning models analyze historical access patterns and regional demand to prefetch frequently requested segments.

[114] Workload is balanced across containers using Kubernetes orchestration, allowing elastic scaling based on user demand.

[115] This layer minimizes round-trip delays, particularly for live and on-demand events in dense urban regions.

[116] D. Latency Optimization Layer

[117] To further reduce end-to-end delay, the system employs AI-driven prefetching to predict and load upcoming segments before playback requests occur.

[118] Transport is managed through congestion-aware protocols such as QUIC and HTTP/3, reducing handshake overhead and improving packet recovery in lossy environments.

[119] The combination of protocol-level optimization and predictive scheduling ensures consistent playback, with target end-to-end latency under 200–300 ms for live streaming scenarios.

[120] E. Integration and Workflow

[121] The four modules operate in a coordinated pipeline:

[122] User requests are redirected to the nearest edge node.

[123] Edge node retrieves base-resolution video from cloud or cache.

[124] AI-based upscaler enhances the video to target resolution.

[125] DRL-powered adaptive delivery engine selects bitrate/resolution dynamically.

[126] Latency optimization mechanisms ensure low-delay delivery to the user device.

[127] This integration of computational intelligence across the pipeline transforms edge nodes from passive content relays into intelligent decision-making agents, capable of optimizing bandwidth, predicting demand, and maintaining high QoE under variable network conditions.

[128] IV. CHALLENGES, GAPS, AND IMPLEMENTATION OUTLOOK

[129] Despite recent advances in edge-based video streaming, several open challenges remain before large-scale deployment of intelligent streaming systems can be realized. Based on the surveyed works and our proposed model, we identify the following gaps and outline a concrete implementation roadmap.

[130] A. Scalability and Real-Time Performance

[131] Most AI-based streaming systems, such as DRL for ABR [1] and CNN-based upscaling [4], exhibit high computational demand, limiting deployment at scale. Current solutions often assume powerful GPUs at the edge, which is impractical in resource-constrained nodes.

[132] Implementation Outlook: Our design will employ containerized deployment (Docker + Kubernetes) to allow elastic scaling. Upscaling will use Real-ESRGAN trained on the DIV2K dataset for image super-resolution, optimized with quantized inference (TensorRT or ONNX runtime) to run efficiently on commodity edge GPUs.

[133] B. Adaptation Under Dynamic Network Conditions

[134] Traditional ABR algorithms fail under rapid fluctuations, and DRL approaches [7], [8] require retraining for new conditions. This creates brittleness when deploying across heterogeneous networks.

[135] Implementation Outlook: We will train a Deep Q-Network (DQN)–based ABR agent using the LIVE-NFLX video QoE dataset, where states include buffer occupancy, throughput history, and latency prediction. Training will use reward functions balancing rebuffering penalty, bitrate stability, and SSIM/PSNR-based visual quality metrics. Online fine-tuning at the edge will allow adaptation to local network conditions without retraining from scratch.

[136] C. Latency Optimization Beyond Caching

[137] While caching reduces round-trip delay [11], [12], unpredictable “flash crowd” events [10] still cause congestion. Current methods lack predictive mechanisms to anticipate demand.

[138] Implementation Outlook: We will implement AI-driven segment prefetching by training a LSTM-based demand predictor on historical access logs. This predictor will pre-load segments into edge caches before request bursts occur, targeting an end-to-end latency budget of 200–300 ms for live streams.

[139] D. Privacy and Security at the Edge

[140] Frameworks like REACT [15] show the trade-off between privacy-preserving analytics and added latency. Most current works lack integration of lightweight security.

[141] Implementation Outlook: Our deployment will incorporate Trusted Execution Environments (TEEs) for GPU resource isolation (building on [13]) and homomorphic encryption for feature vectors exchanged with cloud nodes. This allows secure inference while keeping added latency under 50 ms per segment.

[142] E. Standardization and Interoperability

[143] Surveyed works often target specific protocols (e.g., HTTP/2 [2], QUIC [6]) or proprietary hardware [3]. A lack of standard APIs for AI-inference at the edge limits interoperability.

[144] Implementation Outlook: We propose designing the pipeline with open-source components only: gRPC-based communication for edge–cloud coordination, FFmpeg for transcoding, and PyTorch/TensorFlow models for inference. This ensures portability across heterogeneous edge infrastructures.

[145] V. CONCLUSION

[146] This paper has presented a comprehensive survey of edge-based video streaming techniques, with a focus on AI-powered upscaling, adaptive delivery, and latency reduction. By synthesizing eighteen recent contributions, we identified key advancements in reinforcement learning–based adaptive bitrate control, deep learning–driven super-resolution, and intelligent caching at the edge. A hybrid model was proposed that integrates computational intelligence across four modules, AI-based upscaling, DRL-driven adaptive delivery, predictive edge caching, and latency-aware transport optimization.

[147] The survey and proposed framework together emphasize that edge nodes should evolve from passive caches to intelligent agents capable of real-time decision making, prediction, and optimization. This transition is critical for ensuring scalable, low-latency, and high-quality video delivery in bandwidth-constrained and mobile-first environments.

[148] The current limitation of this work lies in the absence of quantitative validation and benchmarking. While surveyed studies, report promising improvements such as 47% latency reduction, 60% bandwidth savings, and 92% user satisfaction, the proposed architecture requires empirical testing on real datasets and network traces.

[149] Future work will focus on implementing the framework in a containerized edge testbed using open datasets such as LIVE-NFLX and DIV2K. Evaluation will include latency, SSIM/PSNR quality scores, rebuffering rate, and QoE metrics. Further research will also explore privacy-preserving federated reinforcement learning and standardized APIs for edge–cloud coordination, ensuring both scalability and interoperability.

[150] In summary, the integration of edge computing and computational intelligence represents a transformative direction for next-generation streaming. By addressing open challenges in scalability, adaptability, and security, intelligent edge frameworks will enable immersive, high-quality, and responsive video services for diverse global audiences.

[151] REFERENCES

[152] Li, Z., Mao, H., Zhu, J., Jiang, Z., Ghodsi, A., Rexford, J., & Stoica, I. (2020). Deep reinforcement learning for internet congestion control

[153] Xu, M., Yu, H., Pan, J., & Sun, L. (2018). HTTP/2-based low-latency live video streaming.

[154] B. Mengistu, “Deep-Learning Realtime Upsampling Techniques in Video Games,” University of Minnesota Morris Digital Well, 2023.

[155] Lundkvist, F. (2021). Deep upscaling for video streaming: a case evaluation at SVT. KTH Royal Institute of Technology, Stockholm, Sweden.

[156] M. Hu, Z. Luo, A. Pasdar, Y. C. Lee, Y. Zhou, and D. Wu, “Edge-Based Video Analytics: A Survey,” 2023.

[157] K. Nassisid, T. David, and K. Muhammad, “Adaptive Video Streaming over 6G Networks: Buffer Control and User Behavior Analysis,” arXiv preprint, 2024.

[158] M. Choi, A. No, M. Ji, and J. Kim, “Markov Decision Policies for Dynamic Video Delivery in Wireless Caching Networks,” IEEE Trans. Wireless Commun., vol. 18, pp. 5705–5718, 2019.

[159] T. Chen et al., “SODA: An Adaptive Bitrate Controller for Consistent High-Quality Video Streaming,” in Proc. ACM SIGCOMM, 2024.

[160] I. Yeregui et al., “Leveraging 5G Physical Layer Monitoring for Adaptive Remote Rendering in XR Applications,” 2025.

[161] Dharuman et al., “Edge-Optimized HLS/DASH and ML-Based CDN Optimization,” 2023.

[162] STL Partners, “3 Reasons Why Edge Will Change Video Streaming,” STL Partners, 2025.

[163] Muvi, “The Role of Edge Computing in Video Streaming,” Muvi, 2025.

[164] M. Bilal and M. Erbad, “Edge-Based Transcoding and Containerization for Video Streaming,” 2017.

[165] J.-H. Choi et al., “Dynamic Video Delivery Using Deep Reinforcement Learning,” 2025.

[166] A. Ghosh, S. Iyengar, S. Lee, A. Rathore, and V. N. Padmanabhan, “REACT: Streaming Video Analytics on the Edge,” in Proc. 8th ACM/IEEE Conf. on Internet of Things Design and Implementation (IoTDI), 2023.

[167] National Science Foundation, “Adaptive Bitrate Control with Edge-Aware AI,” NSF, 2025.

[168] M. Bilal et al., “FoV-Aware Edge Caching for 360° Video,” 2018.

[169] ElasticEdge, “Kubernetes-Based Elastic Edge Frameworks for Live Analytics,” 2022. 

@article{1712471,
      author = {Hardik Jain, Riya Sagar, Snigdha Vijay, Monisha H. M.},
      title = {A Survey on Efficient Edge-Based Video Streaming: Integrating AI-Powered Upscaling, Adaptive Delivery, and Latency Reduction},
      journal = {Iconic Research And Engineering Journals},
      year = {2025},
      volume = {9},
      number = {6},
      pages = {557-563},
      issn = {2456-8880},
      url = {https://www.irejournals.com/formatedpaper/1712471.pdf},
      abstract = {Video streaming has become one of the dominant drivers of internet traffic, demanding scalable, low-latency, and high-quality delivery solutions. Traditional cloud-based streaming suffers from high latency and bandwidth stress under dynamic network conditions. This survey explores edge-based video streaming architectures that integrate AI-driven upscaling, adaptive bitrate delivery, and latency mitigation mechanisms. We analyze recent advancements in edge caching, reinforcement learning?based adaptive streaming, and AI-powered super-resolution. Further, we propose an architectural model combining edge computing and AI for optimized video delivery. The survey highlights current achievements, limitations, and open challenges, offering guidance for researchers and engineers working on next-generation streaming systems.},
      keywords = {Edge Computing, Adaptive Streaming, Latency Reduction, AI Upscaling, Deep Reinforcement Learning.},
      month = {December},
      doi = {https://doi.org/10.64388/IREV9I6-1712471}
  }