Comportamiento de las crías

The performance was analysed with respect to 11 categories of eye conditions to determine its ability to correctly differentiate eyes into one of 5 levels of eye closure. The 11 categories of eye conditions included various levels of eye closure and gaze direction as explained in section 4.1.2. Figure 6-6 shows the mean error and the mean error magnitude of EC measurement for the 11 eye conditions. For evaluation of the EC measurement system, the frames with different degrees of eye closures (categories 1-5) are of primary interest. However, the frames with different gaze directions (categories 6-11) also provided information on the performance of EC measurement system. An example of the EC measurement system’s performance for a subject (subject 2 in a dark room) under the 11 different eye conditions is presented in Figure 6-7.

Figure 6-6. Performance of EC measurement method for various eye closures and gaze directions. The category 1 to 11 sequentially represent the frames where the eyes are fully open, ¼ closed, ½ closed, ¾ closed, fully closed, and have left gaze, extended-left gaze, right gaze, extended-right gaze, upward gaze, and downward gaze. Apart from frames with closed eyes (category 5), the mean EC error was relatively small in frames with other 4 degrees of eye closures (category 1-4). In most of the gaze direction categories (category 6-11), the mean EC error were positively biased, i.e., the eyes in these frames were estimated to be more closed than they were. The mean error magnitude (mean absolute EC error) shows that the EC measurement was worst in frames where the eyes were closed, followed by frames where the subjects were looking upwards.

Figure 6-7. Example of EC measurement in annotated frames from subject 2 corresponding to 11 categories of eye condition (see section 4.2.1). The images from the top left to the bottom right represent the eye condition categories from 1 to 11 respectively. The EC is measured based on the position of the UELy

(dashed line) relative to the reference height (Ĥ ) and position (UELy_open) represented by the two solid lines.

The mean EC of the left and right eyes for each of these images were 0.03, 0.14, 0.38, 0.57, 0.98, 0.08, 0.03, 0.04, -0.20, -0.19, 0.83, respectively. Reference height (Ĥ₎ and position (UELy_open) Estimated UELy Keys

Figure 6-6 shows the mean EC error for the first 4 levels of eye closure (categories 1-4) was within ±0.125. Therefore, as explained in section 6.2, the EC measurement system can differentiate eyes into one of these 4 levels of eye closure when the subject is looking straight ahead.

However, Figure 6-6 also shows that the mean error of EC measurement was considerably high and negatively biased when the eyes were fully closed (category 5). The poor performance of EC measurement system in majority of frames with fully closed eyes is mainly due to incorrect estimation of the UELy above its true position as explained in section 5.8.2. In the majority of frames with closed eyes, UELy is usually detected at approximately half way between the Ĥ of the eyes, as shown in Figure 6-8, which results in the eyes estimated as half rather than fully closed. Consequently, it is difficult for the EC measurement system to differentiate fully closed eyes in its current state. Nevertheless, if the temporal information about degree of eye closure in preceding frames is known, it should be possible to identify when eyes are fully closed in video data. On average, the EC measurement system performed better in frames with frontal gaze (categories 1-5) then in frames with other gazes (categories 6-11). This is promising as subjects are most likely to be looking straight ahead when they are drowsy or about to have microsleep.

Figure 6-8. An image of closed eyes showing incorrect detection of UELy, resulting in the EC being

incorrectly measured as 0.68.

Analysis of frames with various gaze directions (categories 6-11) showed that EC measurement was particularly poor and positively biased in frames with upward gaze (category 10). This was mainly due to incorrect estimation of UELy_open. During upward gaze, the position of the palpebrae fissure of subjects also moves upward relative to the face. However, the detected fROI does not move. Since, the UELy_open is relative to the position and height of the fROI (see section 5.8.3), the UELy_open is incorrectly estimated and consequently the EC is inaccurately measured. Figure 6-9(a) shows as example of incorrect estimation of UELy_open, which results

in poor measurement of EC. However, in some frames, the fROI also readjusts with the upward gaze, as shown in Figure 6-9(b), resulting in more accurate measurement of the EC.

(a) EC = -0.53 (b) EC = - 0.15

Figure 6-9. Measurement of EC is dependent on the accuracy in estimation of UELy_open. (a) The eyes in this

frame have large EC measurement error due to poor estimation of UELy_open. (b) In this frame, the error in

EC measurement is reduced because fROI moved upward during upward gaze resulting in correct estimation of the UELy_open.

In addition to frames with upward gaze, any displacement of fROI produced large error in EC. For example, although the UELy and Ĥ are correctly estimated in Figure 6-10, displacement of fROI results in incorrect estimation of the UELy_open and consequently a large error in EC.

Figure 6-10. The detected fROI in this frame was lower and smaller than its neighbouring frames. This displacement of the fROI resulted in incorrect estimation of the UELy_open and a large error in EC.