Computer Vision for Supporting Visually Impaired People: A Systematic Review

Copyright


I. INTRODUCTION
Globally around the world, in 2010, the number of people of all ages visually impaired is estimated to be 285 million, of whom 39 million are blind, according to the study of World Health Organization (Global Data on Visual Impairments, 2010).
Regrettably, this percentage is expected to increase in the coming decades. Visual impairment has a significant impact on individuals' quality of life, including their ability to work and develop personal relationships. Almost half (48%) of the visually impaired feel "moderately" or "completely" cut off from people and things around them (Hakobyan, Lumsden, O'Sullivan, & Bartlett, 2013). Visual loss inevitably leads to impaired ability to access information and perform everyday tasks. In today's knowledgeintensive society, information access is increasingly crucial, not just for performing daily activities but also for engaging in education and employment. Also, even the simple tasks around the home can be hazardous if our vision is deteriorating. It's critical for visually impaired people to detect and recognize objects around them, especially in a new environment.
Computer vision is the science that gives the capability to computers to sense visually like humans. Computer vision is concerned with methods for acquiring, processing, analyzing, and of useful information from a single image or a sequence of images. It provides features to see and recognize objects like humans that are very helpful for impaired people (Shapiro & Stockman, 2001). We are very interested in computer vision to support a better quality of life for individuals with disabilities, including visual impairment. We realize and believe that technology can enhance individuals' ability to participate fully in societal activities and live independently.
Many technologies in computer vision have been developed to assist people who are blind or visually impaired. This paper focused on presenting a systematic literature review about different algorithms, devices, and supported computer vision tasks for supporting visionimpaired people.

II. METHODS
A systematic approach for reviewing this literature is chosen. A systematic literature review is a means of identifying, evaluating, and interpreting all available research relevant to a particular research question, or topic area, or phenomenon of interest. Individual studies contributing to a systematic review are called primary studies; a systematic review is a form a secondary study (Kitchenham, 2004).

Research Questions
The research questions (RQ) were specified to keep the review focused. They were designed with the Population, Intervention, Comparison, Outcomes, and Context (PICOC) criteria (Kitchenham, 2004). Table 1 shows the (PICOC) structure of the research questions.

Study Selection
The inclusion and exclusion criteria were used for selecting the primary studies. These criteria are shown in Table 3 below.

Inclusion Criteria Exclusion Criteria
Focused on approach in CV for VI

Studies not written in English
Focused on the supported tasks for visually impaired The outcomes of the articles were not related to VI people

Identification of Papers
Included papers were published between 2016 and 2020. There were three elements to our searches: Keyword searching using the search engines: google scholar, an issue-by-issue manual reading of paper titles in relevant journals and conferences, and the identification of papers using references from included studies. A total of 45 papers were screened, and from that, 13 papers were selected.

Data Extraction and Synthesis
The selected primary studies are extracted to collect the data that contribute to addressing the research questions concerned in this review. The properties were identified through the research questions and used to answer the research questions shown in Table 4. The data extraction is performed in an iterative manner. The data extracted in this review is qualitative data the narrative synthesis method was used.

III. RESULT AND DISCUSSION
In this review we aim to answer the research questions conducted, all the research question will be described one by one. But all the summaries can be seen in appendix Table 5.

Proposed Algorithm in CV Used
There are several techniques used to create CV systems for supporting VIP. In our review, we found six techniques. Those are edge detection, object and obstacle detection, image classification, image segmentation, character recognition, and feature extraction.
For Edge detection, we found two algorithms methods; canny edge detector that is used to detect doors (Sivan & Darsan, 2016) and line segment detector, which is also used to detect doors with an accuracy rate of up to 93.2% for the ImageNet dataset (Talebi & Vafaei, 2018). For objects and obstacle detection, we found three algorithms; first is CNN to recognize the color and sign of traffic got mAP of 96% % ; in another research, CNN is also used to detect objects, but not accurate for multi objects in one scene, so they implemented RCNN , second is YOLOv1 used to detect objects and obstacles and the detection rate is up to 89% for all kind of obstacles , and third is YOLOv3 also used to detect objects, and the mAP is 73.19% (Afif, Ayachi, Pissaloux, Said, & Atri, 2020), and in the other research the accuracy rate is up to 95.19% (Joshi, Yadav, Dutta, & Travieso-Gonzalez, 2020) and 92% (Mahmud, Sourave, Islam, Lin, & Kim, 2020). For image classification, we found two algorithms; the first is KNN to match the descriptor extracted (Elmannai & Elleithy, 2018) and SVM to produce a category label for a scene . For Image segmentation, we found two algorithms as well: K-Means clustering to cluster n extracted points of a particular frame (Elmannai & Elleithy, 2018) and FRRN . For character recognition, we only found optical character recognition (OCR) in three studies, where both studies show high accuracy results (Jiang,Gonnot,Yi,& Saniie,(14)(15)(16)(17) May 2017) (Sivan & Darsan, 2016) (Joshi, Yadav, Dutta, & Travieso-Gonzalez, 2020) and feature extraction using BRISK (Sivan & Darsan, 2016); SURF (Dahiya, Issac, Dutta, Říha, & Kříž, 4-6 July 2018) (Mahmud, Sourave, Islam, Lin, & Kim, 2020)   and ORB   (Elmannai & Elleithy, 2018). The summary can be seen in appendix Table 6.

Devices used
There are many types of equipment as well that were used to build CV systems or applications for supporting VIP. Some of the research are still in the software or application development stage, but other research has reached the prototype development stage. Because there are different stages of research, the tools used are also quite diverse. Studies that proposed a wearable device usually use single board computer, but in a study the researcher use ultrabook laptop to be carried in a backpack , another study also use robot to assist VIP (Mahmud, Sourave, Islam, Lin, & Kim, 2020).
For camera use, there are only two studies that use depth cameras. The depth camera used is the Zed Camera , however, in this study the researcher did not focus on the depth of the information because it would be used in further studies and Astra S Camera (Mahmud, Sourave, Islam, Lin, & Kim, 2020) which was used for path planning. Another researcher used smart glass camera for assisting VIP when shopping  .The complete summary can be seen in appendix Table 7.

Supported Tasks
It's quite rare for researchers to include all the features in one project at once. From 13 eligible papers, we found several categories of supported tasks. Those are sign detection, text detection, object detection, door detection, traffic light detection, object tracking, and navigation.
We found an eligible r3esearch that help VIP to recognize symbols in toilets, pharmacies, and trains (Dahiya, Issac, Dutta, Říha, & Kříž, 4-6 July 2018), while another studies help2ed VIP to recognize signs-based-text, then the text will be converted into sound (Jiang, Gonnot, Yi, & Saniie, 14-17 May 2017) (Sivan & Darsan, 2016). To detect objects, two studies try to help VIP to detect moving and not moving objects in the outdoor area   (Joshi, Yadav, Dutta, & Travieso-Gonzalez, 2020), while another study proposed a system that can detect objects in the indoor area (Sivan & Darsan, 2016 A research uses robots to help VIP to navigate and detect objects with a camera attached to the robot (Mahmud, Sourave, Islam, Lin, & Kim, 2020). Another researcher proposed shopping assistants using smart glasses cameras ), a learning medium for visually impaired children by detecting objects around them by giving a description via voice command .
There are also two other studies that help VIP to detect doors (Sivan & Darsan, 2016) (Talebi & Vafaei, 2018, and another research focus on helping VI to detect the colors of pedestrian signals . The complete summary can be seen in appendix Table 8.

IV. CONCLUSION
The included studies are so diverse that it would not be possible to pool the results from them. However, in the majority of studies, positive effects of the use of computer vision for supporting visually impaired people. These effects included: the detection of obstacles, objects, door and text, traffic lights, sign detections and navigation. The results of this systematic review stress that computer vision really have promising potential for persons who are visually impaired. But the biggest challenge for developers now is to increase the speed of time and improve its accuracy, and we expect the future will have a complete package or solution where blind or vision impaired people will get all the solution together (i.e., map, indoor-outdoor navigation, object recognition, obstacle recognition, person recognition, human crowd behavior, crowd human counting, study / reading, entertainment etc.) in one software and in handheld devices like android. We believe that this will happen in the future, and this paper will help the developer to know the very background in broad models. Talebi, M., & Vafaei, A. (2018). Vision-based entrance detection in outdoor scenes. Multimedia Tools and Applications,77(20), 26219-26238. Zientara, P., Lee, S., Smith, G., Brenner, R., Itti, L., Rosson, M., . . . Narayanan, V. (2017). Third Eye: A Shopping Assistant for the Visually Impaired. Computer, 50(2), 16-24. Appendix: