Introduction

This wiki aims to be a place where work related to my MSc in AI research project will be kept.

This project is looking at how machine learning techniques can be used for object detection when working with ultrasonic data. The use case is chiefly aimed at people with visual impairments: can a model be trained, and deployed to a constrained device which would allow an individual with poor vision to safely navigate a hallway containing obstacles?

The world is becoming a more and more connected place, with over 75 billion connected devices predicted by 2025 [1]. Connected devices are an ecosystem of hardware and software elements which are used to connect to the physical world through the use of onboard sensors as well as connecting to the Internet. Typically, connected devices will have constrained computing power, and a smaller amount of memory than would be typical on larger computing systems. Another common aspect of connected devices is they tend to be battery-powered. The "Internet of Things (IoT)" is often used as an umbrella term to describe a wide range of sensor-enabled devices which are also connected to the Internet.

Due to the restricted memory associated with many of these connected devices, coupled with the potential power issues, there is a need to understand how artificial intelligence (AI) models can be designed to operate at the edge.

Terms like the edge, or edge computing are somewhat vague, and here is a good point to try and wrap some sort of definition around them, and understand where they come from.

As has been discussed, with the increasing development of the IoT, more and more devices are being connected to the Internet, which has resulted in large amounts of data being generated. This has resulted in problems for cloud computing services, which now struggle to support the needs of a more connected society, in terms of things like data processing and data storage. And it is into this gap that edge technologies have emerged. The high-level, one-liner to describe edge computing is a new model of computing for performing calculations at the edge of the network. In other words, the calculations are done closer to the system user, and closer to the source of the data which is being processed [2]. This new model, of course, brings its own challenges, some of which have been introduced above.

This paper is intended to serve a number of functions. First, it will be a review of the main AI models which are common or popular today, with a focus on models used for classification tasks, since object or obstacle detection is a classification task. Second, it will be a review of literature as it relates to object detection, and to machine learning at the edge.

A number of other topics will be discussed throughout this work as they are closely related to the topics of constrained devices. Some of these have already been introduced, such as the Internet of Things, and edge computing. While these topics will be discussed throughout, not a lot of time will be given over to exploring their origins or going into a lot of detail on the research being carried out in the respective areas.

Sources

[1] S. R. Department. “Internet of things (iot) connected devices installed base worldwide from 2015 to 2025.” (2016), [Online]. Available: https://www.statista.com/statistics/471264/iot-number-of-connected-devices-worldwide/ (accessed: 22.10.2022)

[2] K. Cao, Y. Liu, G. Meng, and Q. Sun, “An overview on edge computing research,” IEEE Access, vol. 8, pp. 85 714–85 728, 2020, ISSN: 21693536. DOI: 10.1109/ACCESS.2020.2991734.