Publication

Abstract : The fast development of the applications of the Industrial Internet of Things (IIoT) requires the real-time object detection which has the capability to work effectively in energy-restrained edgefog conditions. Unlike the former YOLO-based and lightweight detectors and RL-based edgefog offloading mechanisms, the proposed RAEED framework combines lightweight detection and training-free adaptive inference offloading to optimize the accuracy, latency and energy consumption jointly in IIoT settings. The deep learning-based detectors are very accurate and capable of handling resource-intensive IIoT deployments, however, their systems are frequently computation- and powerintensive, which is not feasible in resource-constrained IIoT applications. In this effort, the present work suggests that a resource-conscious energy-efficient detector (RAEED) system is proposed which allocates inference resources between the edge and fog nodes dynamically projected by resource availability and network characteristics. The framework integrates a convolutional backbone which is lightweight together with adaptive offloading strategy to trade-off between detection accuracy, latency, and energy consumption. The applications to the COCO 2017 dataset under constrained deployment conditions indicate that RAEED can have 82.4% mAP, 83.1% precision and 81.7% recall, with a latency of 47.6 ms and with a consumption of 0.91 mJ/frame. These findings indicate that IIoT systems can greatly manage to find and identify objects with a much better energy and resource savings.