An Edge–Cloud Cooperative Inference System for Video Analytics with Adaptive Resolution and Early-Exit Routing

Abstract

Video analytics pipelines increasingly span resource-constrained edge devices and elastic cloud backends, motivated by the need to reduce end-to-end latency while maintaining accuracy under bursty visual complexity and fluctuating network conditions. At the same time, modern deep models are over-provisioned for many frames, and naïvely transmitting full-resolution video to the cloud wastes bandwidth and energy. This paper presents an edge--cloud cooperative inference system for streaming video analytics that jointly adapts spatial resolution and performs early-exit routing across a multi-exit neural network, while deciding whether to complete inference on the edge or offload intermediate representations to the cloud. The system treats resolution selection, exit selection, and placement as a coupled control problem under latency, energy, and bandwidth budgets. A unified formulation integrates rate--distortion intuitions for adaptive encoding, uncertainty-calibrated confidence gating for early exits, and placement-aware feature compression for communication efficiency. The design includes an online policy that uses embedding similarity to exploit temporal redundancy, a cost model grounded in operator-level performance measurements, and a distributed execution runtime that supports pipelined micro-batching and backpressure. The paper details optimization strategies spanning constrained stochastic gradient methods, Lagrangian relaxation, and Pareto-style trade-offs between accuracy and resource usage, and provides an evaluation methodology emphasizing trace-driven reproducibility, error decomposition, and stress testing under network variability.