An extensible framework for the deployment and management of computer vision workloads on edge platforms

The challenge in deploying computer vision workloads is the decision as to where the computation of the video streams occurs across the edge-to-cloud continuum. The decisions are based on business, technical and workload variables, which may be dynamic in nature. This thesis proposes a Computer Vision Workload Manager, which will provide an open framework for deploying and managing active workloads across physical and virtualised environments. The benchmarking tests demonstrated that edge devices with higher powered processing capabilities delivered superior performance in terms of MIPS/watt and using virtualisation techniques can also provide increased CPU processing performance by up to 22%. The video analytics workload testing demonstrated the significant difference in processing requirements with different video stream bitrates and how video stream saliency affects processing requirements on edge devices by up to 15%. Using the test results, use cases from two industries were then used to demonstrate how the computer vision workload manager provides power management, load balancing, and deployment of new and prioritised workloads. The computer vision workload manager is designed as a framework to bring together a group of edge devices and cloud instances to deliver computer vision workloads. The ability and flexibility for the system to be extended to ensure the business, technical and workload requirements of different workloads and its ability to use current and emerging edge platform technologies are critical to its successful implementation.