Edge collaborative inference and resource scheduling optimization for optical sensing networks

Abstract

The proliferation of optical sensors (e.g., surveillance cameras) in edge Internet of Things (IoT) environments generates massive video streams, posing critical challenges of high latency, network congestion, and data privacy for cloud-centric deep learning inference. To address this, we propose a novel edge-cloud collaborative inference framework. Our approach unifies model partitioning and task offloading decisions for lightweight vision models. We formulate a joint optimization problem aimed at maximizing the number of served tasks under stringent resource and latency constraints, and devise both offline and online heuristic scheduling algorithms. The online algorithm incorporates a novel dynamic threshold-based admission control mechanism for efficient real-time decision-making. Extensive simulations, based on real-world sensor data profiles and state-of-the-art lightweight vision models (e.g., MobileViT), validate the framework's efficacy. Compared to conventional baselines (FIFO, LBF), our strategy achieves: 1) The task acceptance rate has increased by 15-30%; 2) The uplink data volume for a single task has decreased by approximately 40%; 3) The system throughput has increased by more than 20%, and it strictly adheres to the task deadline. The proposed strategy provides an efficient and practical solution for resource-constrained optical sensing networks, significantly advancing the feasibility of deploying large-scale, real-time intelligent perception at the network edge.