Designing a pervasive eye movement-based system for ALS and paralyzed patients

Digital Alternative Communication for Individuals with Amyotrophic Lateral Sclerosis: What We Have

“…In the set of primary studies, mainly for the face detection/localization and eye detection/localization steps, the use of Computer Vision algorithmic resources from the Open Source Computer Vision Library (OpenCV) [ 81 ] and Dlib [ 82 ] was observed. For the face detection/localization task, for example, the authors mentioned the use of techniques based on the Viola–Jones algorithm or the Haar cascade classifier, available in the OpenCV library, and on Dlib’s built-in facial landmark detector [ 56 , 58 , 60 , 65 , 70 , 74 , 80 ]. Other sources of computational tools for image processing and Computer Vision were also explored in the face detection/localization step: for example, Singh and Singh [ 66 ] and Singh and Singh [ 67 ] used the Viola–Jones algorithm from the MathWorks company.…”

“…For enhancing the pupil (or iris) in an eye image, the robust and versatile technique for detecting circles called the Circular Hough Transform [ 58 , 62 , 63 , 73 , 74 ] is highlighted. Other techniques covering the eye detection/localization task can also be mentioned, such as the Limbus tracking method [ 59 ], gradient-based algorithms [ 60 , 80 ], Machine Learning models based on Hierarchical Temporal Memory [ 61 , 64 ], Erosion with a cross-shaped structure element [ 68 ], segmentation based on thresholding [ 69 ], and the Clustering Method of Unbroken Pixel Lines [ 76 ].…”

“…In the set of studies where the purpose was to develop a Human–Computer Interaction strategy based on gaze direction, it is evident that the authors followed the workflow, to identify the pupil/iris in the images in advance, using varied methods such as Circular Hough Transform [ 58 , 62 , 63 , 73 , 74 ] or gradient-based algorithm [ 60 , 80 ], and that they extracted values related to the coordinates that served as input to mathematical models (called Geometrical Dependencies in this work) that calculated and identified the gaze direction [ 58 , 60 , 62 , 63 , 73 , 74 , 78 , 79 , 80 ]. The authors Eom et al [ 56 ] and Yildiz et al [ 62 ] used Geometrical Dependencies to train and create Machine Learning models, using neural networks and the K-Nearest Neighbor algorithm, respectively.…”

“…Abe et al [ 59 ] presented the average error in two perspectives of eye-gaze detection: vertical detection (0.56 ) and horizontal detection (1.09 ). The authors Rahnama-ye-Moqaddam and Vahdat-Nejad [ 60 ] reported an average error rate of 5.68% and, now looking at the best error rate obtained by the system, Yildiz et al [ 62 ] presented 0.98% as a result. All the computational techniques explored are identified in Table 4 .…”

“…Based on the Video Acquisition step presented in Figure 3 , 76% of the primary studies [ 56 , 57 , 58 , 59 , 60 , 61 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 74 , 76 , 78 , 79 , 80 ] performed computational experiments using devices for image collection placed on a table or integrated into the computer itself, as in the case of notebooks or smartphones with integrated cameras, which characterizes a Human–Computer Interaction approach where users are free of devices on their body. Alternatively, 24% of the studies explored a Human–Computer Interaction approach where the prototype for image collection, the camera, was mounted on the user’s head [ 62 , 63 , 64 , 73 , 75 , 77 ].…”