基於眼球注視偵測技術的自閉症孩童早期快篩系統

A Rapid Screening System for Early Autism Children Based on Eye Gaze Detection Techinques

蔡東昇(Dong-Sheng Tsai)

指導教授 ： 繆紹綱

中原大學/電機資訊學院/電子工程研究所/碩士(2016年)

https://doi.org/10.6840/cycu201600722

摘要

眼球注視偵測的技術對於現今智慧型裝置普遍發展的社會有許多不同的用途，例如有些智慧型手機能偵測使用者在觀看影片時眼睛是否有停留在螢幕上，以此來決定正在觀看中的影片要暫停或者是播放。另外，這項技術也能運用在廣告曝光率的統計、及時對需要協助的孩童伸出援手以提升教育學習的效率、幫助漸凍人患者與外界溝通以及作為測謊器對於犯罪的偵查等等。
眼球注視偵測的技術可以分成接觸式和非接觸式兩大類。接觸式的方式有眼電圖法和搜尋線圈法，非接觸式的方式有瞳位追蹤法、紅外線眼動圖法。上述的這些方法都會對人的眼睛造成傷害，尤其是接觸式的方法。在進行接觸式的眼球注視偵測時須先將眼睛局部麻醉才能進行測試，而非接觸式的方式大部分都要使用到紅外線LED光源來使瞳孔影像更加明顯，但也會造成眼睛的乾澀和不舒服感，而且長期使用紅外線LED光源照射眼睛也會對眼睛造成傷害。
為了避免上述問題，我們參考了在異色邊界處理法中使用的一般光線及常見的攝影鏡頭來進行實驗。但由於一般的光線相較於紅外線LED光源難以凸顯眼睛特徵，因此對後續實驗步驟的處理上有一定的困難在。在這部份我們參考異色邊界處理法中所提到的三點測圓的方法，找出瞳孔的位置並採用現有的臉部偵測技術使其系統能在不使用支架固定頭的情況下偵測出眼睛的位置，並估算出瞳孔座標在螢幕上的注視點為何。
本論文的另一個重點是探討如何將這套系統運用到特殊教育上。在本論文中將設計出結合注視點系統的故事劇情來進行實驗。透過此種實驗來快速篩檢出有自閉症傾向的孩童以減輕專業人士的負擔。由於我們僅需配備攝影機鏡頭的個人電腦即可進行使用，可方便在家裡進行快速篩檢。在進行快篩檢測時，我們會在對小孩較少吸引力的空間裡對小孩子進行測試，利用說故事的方式吸引小孩子的注意力並與其進行互動。在故事過程中會隨著故事情節設計有關辨識人物情緒的問題，利用自閉症孩童在辨識情緒上的困難來檢測孩童是否有自閉症的傾向。
實驗結果顯示，我們的系統能抓取到雙眼瞳孔注視中間時的成功率平均能達到94%，當眼睛注視在上方及左右這些方向時我們也能平均達到93%左右的成功率，但當我們在注視下方時，由於眼皮的遮蓋所以在這部分的成功率僅能平均達到89%。在注視點估測方面，注視點估測的結果和我們理想中的注視座標會有些差距，雖然無法準確地知道實驗對象是注視在螢幕上的確切位置，但依舊能分辨出實驗對象所注視的方向是在螢幕上的那些區域。因此我們利用能分辨出實驗對象所注視的區域進行模擬實驗，在模擬的過程中我們探討使用兩種不同的分界線所分隔的螢幕畫面和在各區域上的三種不同大小的目標影像進行實驗，模擬的結果發現使用兩種不同的分界線會對模擬結果造成些微差距，主要是所注視的目標影像大小會影響實驗結果，實驗結果顯示大型目標影像所造成的誤判最多，中型和小型的目標影像誤判較少。基於注視影像要清楚的前提下，本實驗建議選用中型影像當作我們的模擬注視影像。
未來我們將嘗試改良本系統，包括減少注視點估測誤差以及系統運算時間，再將此系統應用到教育和商業等更多地方。

關鍵字

人臉偵測 ； 眼球注視 ； 異色邊界地帶 ； 注視點估測 ； 自閉症

並列摘要

As the popularity of smart devices gains rapidly nowadays, eye gaze detection technology can be used in various applications. For example, some smartphones can detect if the eyes of a user stay on the screen while watching a video, and then decide to pause or play the video. This technology is also applicable for analyzing the advertisement exposure, enhancing the efficiency of teaching by immediately helping children facing learning difficulties, being a communication tool between ALS (Amyotrophic Lateral Sclerosis) patients and others, and serving as a polygraph in criminal investigation.
Eye gaze detection methods fall into two major categories, contact and non-contact. The electrooculography (EOG) system and the scleral search coil method are typical examples of contact type, while the non-contact type includes a pupil position tracking method and infrared oculography. All the methods mentioned above can cause some damage to human eyes, especially the contact type. When we test a contact-type eye gaze system, the eyes of a subject must be under local anesthesia. Furthermore, most of the non-contact type approaches need to use an infrared LED source to make the pupil image more obvious, which can cause eye dryness and discomfort. Besides, the long-term use of infrared LED source would be harmful to the eyes.
In order to avoid the problems above, we consider the limbus-based approach with an ordinary light source and a common photographic lens in our experiment. However, as compared to the use of infrared LED, it is harder to highlight the eye characteristic by the ordinary light source, presenting some difficulties in subsequent experiments. Thus, we adopt the three-point circle detection method from the limbus-based approach to find the position of the pupil and incorporate an existing face detection technique so that the system can detect eyes without the use of a head supporting stand, and estimate the eye gaze point on the monitor.
In this thesis, we also focus on how to apply the system to special education, where we design a story to detect the emotion reaction of a child by employing the eye gaze system in our experiments. Through this rapid screening test for the children with autistic tendency, we may be able to reduce the burden of professionals. Since we only need to equip with camera lens and PC, a rapid screening can be carried out conveniently at home and then send the screening results to the professionals for further evaluation. We conduct the rapid screening test in a space without too much distraction for children under test, as we tell a story to attract the attention from them and interact with them. As the story goes, the subject will be asked to identify the correct emotion expression of the characters in the story. The difficulty in identifying emotion expression is exploited to detect the children with autistic tendency.
The experimental results show that the system can capture the pupils at about 94% success rate when the eyes gaze at the screen center. When the eyes gaze at left, right and up positions of a screen, the success rate is about 93%, while it is only 89% or so due to the eyelid cover when we look downward. In the gazing point estimation, location errors exist between the gazing point estimated and the exact gazing point. Although we cannot identify the exact gazing point on the screen, we are still able to distinguish the correct gazing area with large enough size on the screen. We simulate the rapid screening test by gazing at three circle objects with different sizes, where the object represents a face of the character with certain motion expression in the story. In the simulation, the two objects with the same size are separated by two different boundary lines. Experimental results show that there are little differences with different boundary lines to separate the objects. In fact, the experimental results are mainly affected by the object size. Specifically, large objects result in higher classification error, while medium and small objects have a lower chance of misclassification. Since a larger object can reflect more details of facial expression, we choose the medium size object for our rapid screening system.
In the future, we will try to improve the proposed system by reducing the estimation error of gaze point and the operation time of the system and apply the system to more areas, including education and business.

並列關鍵字

Face detection ； eye gaze ； limbus-based approach ； gazing point estimation ； autism.

參考文獻

[15] 江知遠，以異色邊界地帶瞳孔追蹤處理法進行注視落點偵測，中原大學通訊工程碩士學位學程碩士論文，2013。

[29] 財團法人中華民國自閉症基金會，

[2] 簡郁芩，以眼動型態探討背景知識對詞彙辨識的影響，國立中央大學學習與教學研究所碩士論文，2006。

[3] Eckhard H. Hess and James M. Polt, “Pupil Size in Relation to Mental Activity during Simple Problem-Solving,” Science 13, Vol. 143, Issue 3611, pp. 1190-1192, Mar. 1964.

[10] Andreas Bulling, Hans Gellersen, and Gerhard Troster, “Eye Movement Analysis for Activity Recognition Using Electrooculography,” IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 33, No.4, pp. 741-755, Apr. 2011.