Abstract
Computer vision applications are increasingly used on mobile Internet-of-Things (IoT) devices such as drones. They provide real-time support in disaster/incident response or crowd protest management scenarios by e.g., counting human/vehicles, or recognizing faces/objects. However, deployment of such applications for real-time video analytics at geo-distributed areas presents new challenges in processing intensive media-rich data to meet users' Quality of Experience (QoE) expectations, due to limited computing power on the devices. In this paper, we present a novel policy-based decision computation offloading scheme that not only facilitates trade-offs in performance vs. cost, but also aids in offloading decision to either an Edge, Cloud or Function-Centric Computing resource architecture for real-time video analytics. To evaluate our offloading scheme, we decompose an existing computer vision pipeline for object/motion detection and object classification into a chain of container-based micro-service functions that communicate via a RESTful API. We evaluate the performance of our scheme on a realistic geo-distributed edge/core cloud testbed using different policies and computing architectures. Results show how our scheme utilizes state-of-the-art computation offloading techniques to Pareto-optimally trade-off performance (i.e., frames-per-second) vs. cost factors (using Amazon Web Services Lambda pricing) during real-time drone video analytics, and thus fosters effective environmental situational awareness.