CAPÍTULO 5- OTRAS TAREAS DESARROLLADAS EN LA EMPRESA
5.5. Conclusiones
The Blob Analysis method calculates statistics for labelled regions in the binary image and returns the area of each element. With ‘rough’ searching, the eye sclera area can be envisaged as connected blobs. While approximating a blob shape by use of an ellipse, the area was calculated (Horn, 1986) (see Appendix B1 for eye sclera area calculation equation). The Blob analysis technique in MATLAB has standard blocks for area calculation. The best match for eye sclera region detection is the ellipse shape.
Parameters of the ellipse such as shape, aspect of short and long axes, and the size of area are used to decide the area of eye sclera section. The captured eye blink in three different stages is shown in Figure 5-16 (a), (b), (c).
Figure 5-16 (a): Full opened eye sclera region, image from left- (i) captured eye sclera image using green illumination system, (ii) Blob analysis tracking image.
Figure 5-16 (b): half way closed eye, eye sclera area reduced in binary image
Figure 5-16 (c): captured 3/4 of closed eye
In the Blob analysis technique (the technique calculates statistics for labeled regions in a binary image) for tracking and measuring the eye sclera area, the major and minor axis lengths have been calculated to design a common threshold to detect blink for different people having different sizes of eye sclera region. The procedure is as follows:
• Estimate common threshold level for measuring the eye sclera area to detect blink duration.
• Threshold level needs to readjust for different participants with reference to their eye sclera area.
This PhD research considered the displacement and speed of blobs for behavioural state estimation. The eye region consists of the upper and lower eyelids and eyelashes (for blob detection, see Figure 5-17 (a)). In the detection process, the eye region has been divided into upper and lower portions. The intensity distribution is changing when the upper and lower portions of the eyelids close and open during blink. When the eye lids are closing the sclera area is reduced, and when the eye lids
are opening the area increases. This concept is used to calculate the eye blink duration. The blob analysis starts to measure blink duration when the eye lids shuts to half - to full close - to half open. This procedure is considered as a complete eye blink. The major and minor axis lengths of the ellipse change according to eye sclera region changes as illustrated in Figure 5-17 (a), (b) and (c). (Xleft, Yleft - Xright, Yright
) is the major axis coordinates and during the eye close this axis will reduce. The minor axis coordinates are (Xtop, Ytop – Xbottom, Ybottom) and will reduce during the eye
close. These lengths change within the range {Xi, Yi; i= left, top, right, bottom} from
the full eye open to full eye close, and the range of the axis changes is given in the following equation.
I(x,y)=
{
I(x,y)x
left≤
x
≤
x
right,y
top≤
y
≤
y
bottom}
(5-8)(a) (b) (c)
Figure 5-17: (a) The Blob detection model of eye sclera region detect using ellipse method, (b) Half eye close position, (c) quarter of eye close position.
Figure 5-17 (a), (b) and (c) illustrate the three positions of eye closure. The next position after (c) is full closed image and it appears as a blank image in detection. To measure the accuracy of the blink detection system, the total blinks during the 40 minute simulator test were manually counted for randomly selected ten participants and compared with the total blinks counted from the system. Figure 5-18 (a) shows
an example of captured video file conversion to frame by frame image file for manual blink count. Figure 5-18(b) shows the actual eye blink count and the eye blink count from the system. The average of the actual blink count for ten participants was 397.6 and the average eye blink count from the system is 388.3. The error was 2.3% and the accuracy of the detection system was 97.7%.
Figure 5-18 (a): Frame by frame eye images used for the blink count.
In the validation of a common threshold pixel value for the eye sclera region (full eye open), the Simulink model was designed to measure different eye sclera regions in the half eye open position for all participants. The average pixel area of the half eye position was quantified to select the common threshold for all participants. The Figure 5-19 illustrates the Simulink model designed to measure the eye sclera area for different sizes of sclera regions (i.e., Figure 5-16 (a), (b) and (c) shows binary image sclera region selected to quantify the pixel area from quarter (1/4) eye open to full eye open. The pixel area change 5-16(a) 3410 pixels (full eye open), 5-16(b) 2928 pixels (half eye open) and figure 5-16 (c) 2401 pixels (quarter eye open). The Blob detection ellipse axis change, according to the above positions [(a), (b) and (c)], is [(67.73, 115.2), (53.66, 85.07), (50.14, 65.13)]. Figure 5-20 shows the calculated eye sclera area for eighteen participants when their eyes are half open. The mean eye sclera value for all participants = 2799 pixels and the SD of sclera area is for all participants = 394.3 pixels. The eye sclera area in full eye open position for all participants is averaged to make a common threshold value for eye blink detection. The common threshold value is 3000 pixels, and this value was calculated by adding the participants’ mean sclera value to participants’ SD of sclera area (2799+394= 3193 ≈ 3000 pixels).
Figure 5-19: The Simulink model designed to measure different size of eye sclera regions.
Figure 5-20: Eye sclera areas for eighteen participants for half eye open.