Multimedia Systems Компьютерные сети и коммуникации
FADE: feature-aware deterministic pruning for real-time edge video analytics
Аннотация
Multi-stream pedestrian surveillance must adapt to changing crowd conditions under limited edge resources. Our observations motivate jointly adapting resolution, encoding bitrate, and analytics frame rate to maintain detection and tracking quality. However, content estimation and configuration search introduce additional overhead. We propose FADE (Feature-Aware Deterministic Pruning), a lightweight framework for content-aware configuration adaptation in edge video analytics. FADE reuses decoder-produced grayscale images to estimate foreground texture entropy within detected regions, and downstream detection and tracking results to estimate object scale and motion intensity. This approach avoids an additional online perception network while keeping feature-estimation overhead low. Offline-calibrated models relate these features to configuration requirements. FADE first narrows the configuration space using content-dependent constraints, then screens candidates against a predicted detection-accuracy threshold. An online scheduler jointly selects configurations from the reduced sets and allocates shared bandwidth. Lyapunov optimization and bandwidth-price coordination balance energy consumption and predicted task utility under long-term average modeled-latency constraints. Prototype experiments show that two-stage pruning reduces the average candidate count by 75.2% and configuration-search time from 11.5 to 2.3 ms. Across three fixed-bandwidth multistream experiments, detection accuracy ranges from 93.3% to 94.1%. At 80 Mbps, FADE supports 55 concurrent streams with an end-to-end latency of approximately 53.2 ms. These results demonstrate that reusing existing pipeline outputs and pruning candidate configurations reduce online adaptation overhead while supporting efficient multistream resource coordination.