An Edge Intelligence Framework for Resource Constrained Community Area Network

Edge intelligence, Artificial Intelligence (AI) on the edge can have a significant impact on modern Community Area Network (CAN). This paper proposes an edge intelligence method that utilizes deep learning, object detection, and multi-label multi-classification to perform monitoring and actuation tasks without resorting to high-end edge servers. The proposed method contains a resource-constrained node as an edge device. For the edge server, it utilizes a special-purpose ASIC (Intel's Movidius) interfaced with a node-level edge device. To further the idea of limited bandwidth availability in CAN, pseudo D2D communication is employed. SSD-MobileNet and customized multi-label-multi-classification based GoogLeNet models are hosted on the edge server, The proposed methodology can achieve about 5.26 FPS for complete bi-directional communication.