Online Outcome-Driven Configuration Selection With Adaptive Grouping for Live Video Streams
- 5 May 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. PP (15361233) , 1-15
Abstract
Optimizing live video stream configurations, such as resolution, is essential for balancing inference accuracy and bitrate, thereby reducing data volume and conserving network resources in intelligent video analytics. Existing systems predominantly rely on offline or periodic strategies using historical data to determine optimal configurations. However, these methods may fall short in managing bitrate, which is highly sensitive to dynamic variations in video content and can benefit from real-time adjustments. To address these limitations, we propose the O $_{2}$ MEGA (Online Outcome-driven Grouping Configuration Adaptation) framework. Motivated by unique experimental observations, O $_{2}$ MEGA adapts configurations in an “outcome-driven” manner by directly utilizing real-time online bitrate outcomes from encoded video streams. This approach allows efficient and lightweight integration of outcome feedback into configuration decisions without incurring significant analysis overhead that could negate the benefits of adaptation. Moreover, O $_{2}$ MEGA introduces a dual grouping mechanism that enables cross-camera collaboration on configuration sharing, expediting configuration selection in densely deployed camera networks. Unlike existing methods that rely solely on spatio-temporal correlations, our framework groups cameras based on real-time outcome feedback, thus adaptively optimizing group-level performance online. We provide theoretical guarantees to ensure the interpretability and effectiveness of the framework. Experimental results on a testbed with four NVIDIA development kits and an edge server demonstrate up to 6.8% accuracy improvement or 13.2% data volume reduction compared to the state-of-the-art under four typical vision queries.
Keywords
Funding Information
- National Natural Science Foundation of China (625B2163)
- RGC GRF (14202923)
This publication has 50 references indexed in Scilit:
- Efficient Online DNN Inference with Continuous Learning in Edge Computing2024
- End-Edge Coordinated Joint Encoding and Neural Enhancement for Low-Light Video Analytics2023
- Maxim: DRL-Based Cross-Camera Streaming Configuration for Real-Time Video Analytics2022
- CASVA: Configuration-Adaptive Streaming for Live Video Analytics2022
- Bandit Algorithms2020
- SurveilEdge: Real-time Video Query based on Collaborative Cloud-Edge Deep Learning2020
- DeepDecision: A Mobile Deep Learning Framework for Edge Video Analytics2018
- YOLO9000: Better, Faster, Stronger2017
- A large-scale benchmark dataset for event recognition in surveillance video2011
- Rate control of MPEG video coding and recording by rate-quantization modelingIEEE Transactions on Circuits and Systems for Video Technology, 1996