東華大學圖書館 |

Design of a Vision Based Assistive System for Visually Impaired Persons.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Design of a Vision Based Assistive System for Visually Impaired Persons./
作者:	Zientara, Peter Arnold.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2021,
面頁冊數:	103 p.
附註:	Source: Dissertations Abstracts International, Volume: 83-03, Section: B.
Contained By:	Dissertations Abstracts International83-03B.
標題:	Visual impairment. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28841745
ISBN:	9798460448067

Design of a Vision Based Assistive System for Visually Impaired Persons.
Zientara, Peter Arnold.

Design of a Vision Based Assistive System for Visually Impaired Persons. - Ann Arbor : ProQuest Dissertations & Theses, 2021 - 103 p.

Source: Dissertations Abstracts International, Volume: 83-03, Section: B.

Thesis (Ph.D.)--The Pennsylvania State University, 2021.

This item must not be sold to any third party vendors.

Globally, more than 285 million people suffer from some sort of visual impairment. To those with visual impairment everyday tasks can pose a challenge. Tasks including picking up a dropped object, navigating to a destination, and grocery shopping can be difficult. While there are many potential solutions to these common tasks, most of them rely on help from others to accomplish them, thus limiting the freedom a visually impaired person can experience in their day-to-day lives. Advances in all areas of technology including cameras, mobile devices, algorithms, and connectivity give rise to the question "Can a computer system be made that is able to assist with these tasks?". This dissertation presents an assistive visual device that combines wearable cameras, hardware accelerators, and algorithms that enables users with limited or no sight to select products from grocery shelves. In this system, both algorithms and hardware are designed to leverage the interactions with the shopper to accomplish this task. Additionally, through optimization of such a system including its algorithms, hardware acceleration, and network communication, this dissertation explores how to make such a visual assistive system have sufficient performance and responsiveness as well as the development of user feedback mechanisms needed for such a system to be used in real-world scenarios. Recent advances in deep neural networks (DNNs) have produced enormous gains, especially in the computer vision domain. While the push for ever-increasing accuracy has grown these networks to dozens of layers and billions of operations, there has been a simultaneous push to harness the power of DNNs on heavily constrained embedded and mobile platforms. This has led to approaches that trade accuracy for efficiency and has fueled the investigation of FPGAs as a means to directly embody these limited precision models while still being able to adapt to the rapid pace of DNN algorithm development. However, compared to the dominant approach for neural network inference in non-embedded domains, i.e., GPUs, FPGAs currently exhibit substantial drawbacks in programmability, especially in terms of multi-device scenarios. This dissertation also introduces an automated tool flow that goes from DNN definition to embedded system implementation with FPGA accelerated inference and demonstrates the capabilities of the proposed framework in distributing the execution of these neural networks among a set of connected devices. We explore, through simulation, the design space of computational capability and connectivity levels expected across different collaborative embedded system scenarios, such as drone swarms. We then use our flow to perform offload in a real multi-FPGA scenario. We show that, when provided with information about current network and ancillary compute availability, distributing the workload of a neural network can result in speedups up to the limit of available compute, providing the greatest benefits for networks that did not originally target mobile devices. In total, this dissertation presents a) The design and development of a visual assist platform that incorporates a human-in-the-loop feedback mechanism, b) Optimizations of such a system where there are heterogeneous compute elements (CPUs, GPUs, FPGAs, ASICs) that affect the system's responsiveness, performance, power, and accuracy, and c) The creation of a design automation framework for distributed intelligence on an FPGA based system on chip.

ISBN: 9798460448067Subjects--Topical Terms:

3681233
Visual impairment.
Subjects--Index Terms:

Visual impairment

Design of a Vision Based Assistive System for Visually Impaired Persons.
LDR:04713nmm a2200373 4500 001 2342913
005 20220415153039.5
008 241004s2021 ||||||||||||||||| ||eng d
020 $a 9798460448067
035 $a (MiAaPQ)AAI28841745
035 $a (MiAaPQ)PennState_24751paz117
035 $a AAI28841745
040 $a MiAaPQ $c MiAaPQ
100 1 $a Zientara, Peter Arnold. $3 3681315
245 1 0 $a Design of a Vision Based Assistive System for Visually Impaired Persons.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2021
300 $a 103 p.
500 $a Source: Dissertations Abstracts International, Volume: 83-03, Section: B.
500 $a Advisor: Narayanan, Vijaykrishnan;Sampson, John.
502 $a Thesis (Ph.D.)--The Pennsylvania State University, 2021.
506 $a This item must not be sold to any third party vendors.
520 $a Globally, more than 285 million people suffer from some sort of visual impairment. To those with visual impairment everyday tasks can pose a challenge. Tasks including picking up a dropped object, navigating to a destination, and grocery shopping can be difficult. While there are many potential solutions to these common tasks, most of them rely on help from others to accomplish them, thus limiting the freedom a visually impaired person can experience in their day-to-day lives. Advances in all areas of technology including cameras, mobile devices, algorithms, and connectivity give rise to the question "Can a computer system be made that is able to assist with these tasks?". This dissertation presents an assistive visual device that combines wearable cameras, hardware accelerators, and algorithms that enables users with limited or no sight to select products from grocery shelves. In this system, both algorithms and hardware are designed to leverage the interactions with the shopper to accomplish this task. Additionally, through optimization of such a system including its algorithms, hardware acceleration, and network communication, this dissertation explores how to make such a visual assistive system have sufficient performance and responsiveness as well as the development of user feedback mechanisms needed for such a system to be used in real-world scenarios. Recent advances in deep neural networks (DNNs) have produced enormous gains, especially in the computer vision domain. While the push for ever-increasing accuracy has grown these networks to dozens of layers and billions of operations, there has been a simultaneous push to harness the power of DNNs on heavily constrained embedded and mobile platforms. This has led to approaches that trade accuracy for efficiency and has fueled the investigation of FPGAs as a means to directly embody these limited precision models while still being able to adapt to the rapid pace of DNN algorithm development. However, compared to the dominant approach for neural network inference in non-embedded domains, i.e., GPUs, FPGAs currently exhibit substantial drawbacks in programmability, especially in terms of multi-device scenarios. This dissertation also introduces an automated tool flow that goes from DNN definition to embedded system implementation with FPGA accelerated inference and demonstrates the capabilities of the proposed framework in distributing the execution of these neural networks among a set of connected devices. We explore, through simulation, the design space of computational capability and connectivity levels expected across different collaborative embedded system scenarios, such as drone swarms. We then use our flow to perform offload in a real multi-FPGA scenario. We show that, when provided with information about current network and ancillary compute availability, distributing the workload of a neural network can result in speedups up to the limit of available compute, providing the greatest benefits for networks that did not originally target mobile devices. In total, this dissertation presents a) The design and development of a visual assist platform that incorporates a human-in-the-loop feedback mechanism, b) Optimizations of such a system where there are heterogeneous compute elements (CPUs, GPUs, FPGAs, ASICs) that affect the system's responsiveness, performance, power, and accuracy, and c) The creation of a design automation framework for distributed intelligence on an FPGA based system on chip.
590 $a School code: 0176.
650 4 $a Visual impairment. $3 3681233
650 4 $a Software. $2 gtt. $3 619355
650 4 $a Cellular telephones. $3 607843
650 4 $a Sensors. $3 3549539
650 4 $a Design. $3 518875
650 4 $a Computer science. $3 523869
650 4 $a Artificial intelligence. $3 516317
650 4 $a Disability studies. $3 543687
653 $a Visual impairment
653 $a Assistive devices
653 $a Neural networks
690 $a 0389
690 $a 0984
690 $a 0201
690 $a 0800
710 2 $a The Pennsylvania State University. $3 699896
773 0 $t Dissertations Abstracts International $g 83-03B.
790 $a 0176
791 $a Ph.D.
792 $a 2021
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28841745