A Distributed Framework to Orchestrate Video Analytics Applications

TL;DR

This paper introduces a distributed framework for orchestrating video analytics in IoT-based smart doorbells, enhancing transparency, reusability, and flexibility across Edge and Cloud resources, with a comparative evaluation of performance metrics.

Contribution

It proposes a novel distributed architecture for video analytics that improves modularity and transparency, addressing limitations of existing proprietary and monolithic solutions.

Findings

AWS-based approach achieves high object detection accuracy

Low memory and CPU usage compared to state-of-the-art

Higher latency observed in the proposed framework

Abstract

The concept of the Internet of Things (IoT) is a reality now. This paradigm shift has caught everyones attention in a large class of applications, including IoT-based video analytics using smart doorbells. Due to its growing application segments, various efforts exist in scientific literature and many video-based doorbell solutions are commercially available in the market. However, contemporary offerings are bespoke, offering limited composability and reusability of a smart doorbell framework. Second, they are monolithic and proprietary, which means that the implementation details remain hidden from the users. We believe that a transparent design can greatly aid in the development of a smart doorbell, enabling its use in multiple application domains. To address the above-mentioned challenges, we propose a distributed framework to orchestrate video analytics across Edge and Cloud resources. We investigate trade-offs in the distribution of different software components over a bespoke/full system, where components over Edge and Cloud are treated generically. This paper evaluates the proposed framework as well as the state-of-the-art models and presents comparative analysis of them on various metrics (such as overall model accuracy, latency, memory, and CPU usage). The evaluation result demonstrates our intuition very well, showcasing that the AWS-based approach exhibits reasonably high object-detection accuracy, low memory, and CPU usage when compared to the state-of-the-art approaches, but high latency.