An Architecture for Distributed Video Stream Processing in IoMT Systems

In Internet of Multimedia Things (IoMT) systems, Internet cameras installed in buildings and streets are major sources of sensing data. From these large-scale video streams, it is possible to infer various information providing the current status of the monitored environments. Some events of interest that have occurred in these observed locations produce insights that might demand near real-time responses from the system. In this context, the event processing depends on data freshness, and computation time, otherwise, the processing results and activities become less valuable or even worthless. An encouraging plan to support the computational demand for latency-sensitive applications of largely geo-distributed systems is applying Edge Computing resources to perform the video stream processing stages. However, some of these stages use deep learning methods for the detection and identification of objects of interest, which are voracious consumers of computational resources. To address these issues, this work proposes an architecture to distribute the video stream processing stages in multiple tasks running on different edge nodes, reducing network overhead and consequent delays. The Multilevel Information Fusion Edge Architecture (MELINDA) encapsulates the data analytics algorithms provided by machine learning methods in different types of processing tasks organized by multiple data-abstraction levels. This distribution strategy, combined with the new category of Edge AI hardware specifically designed to develop smart systems, is a promising approach to address the resource limitations of edge devices.