A Distributed Framework to Demonstrate Video Analytics across Cloud and Edge

Abstract

The Artificial Intelligence of Things (AIoT) is a fast-expanding field of Machine learning applications and technologies that comprises algorithms, hardware, and software for analysing sensor data (i.e., temperature vibration, audio, and vision data). AI of things can be roughly classified into three main domains- in order of increasing compute, these are vibration, voice, and vision. Vibration includes systems that do predictive maintenance localized to the specific motor. Voice includes microphones that detect voice keywords and recognise speech. Vision includes industrial systems recognizing objects to be able to sort items and spot defects or systems detecting human precedence, or systems identifying faces to unlock devices. All of these scenarios have distinct workload, performance needs and demands scalable solutions. The focus of this paper is on Vision applications. Concept of Computer vision - computers interpret images and videos has been around since the 1960s. Because of computer vision, many technological improvements are achievable. IoT devices are getting cheaper, smaller and becoming computationally and energy efficient with each iteration. A few essential features must be addressed before creating any IoT-based video analytics application, such as cost-effectiveness, extensive usage, scalable design, precise scene detection, and framework re-usability. One such application domain for video analytics in which several commercial options are available in the consumer market is video-based Smart Doorbell for the home. The current offerings, on the other hand, are expensive, inflexible, and exclusive, meaning that the consumer is unaware of the implementation specifics. There will also be a trade-off between the product’s precision and portability. To solve the observed issues, we propose a video analytics framework with the use case of a Smart Doorbell. The suggested architecture leverages AWS cloud services as a basic platform, and the system was built on a low-cost Raspberry Pi to fulfil the price affordability limitation. The Smart Doorbell will be able to distinguish familiar and unfamiliar people with high accuracy. Along with this, additional functionalities include animal/pet detection, harmful weapon detection, noteworthy vehicle detection, logo detection for delivery companies, and package detection. Finally, as a part of our research, we wanted to collate the state-of-the-art video analytics approaches with Privacy & Security, Energy usage and Opacity. Along with that we also wanted to compare them on metrics such as hardware & software cost, resource usage, accuracy and latency and tried recommending the best state-of-the-art approach and device based on the use case.