- 學位論文
以異色邊界地帶瞳孔追蹤處理法進行注視落點偵測
Eye gaze detection using a limbus-based pupil tracking method
摘要
眼球追蹤有許多用途，例如有些智慧型手機可偵測觀看影片時觀看者的眼睛是否有停留在螢幕上，若為否就自動暫停影片的播放。此外，也有平板電腦廠商在平板上加裝一個紅外線LED用來偵測眼球位置，記錄使用者的視線落點數據，廣告公司也可利用這些數據去做廣告的效益分析。這些裝置不加裝LED就只能偵測使用者是否有在看螢幕，無法進行落點偵測，加裝LED後比較能估測出落點位置，但LED會讓使用者眼睛乾澀甚至受到傷害。 要追蹤眼球有許多方法，大致可分成接觸式和非接觸式兩類。接觸式的如眼電圖法、搜尋線圈法，非接觸式的有瞳位追蹤法及紅外線眼動圖法。很不幸的這些方法都會直接或間接的對眼睛造成物理上的傷害，在接觸式的方式中以搜尋線圈法最具侵入性，也必須先將眼睛做局部麻醉，而非接觸式的方式中，大多使用到紅外線LED，這可以使眼睛瞳孔特徵更明顯，但也會造成眼睛乾澀不舒服。 為了避免眼球追蹤時使用侵入性設備及紅外線LED燈對眼睛的傷害，可以利用一般日光燈自然照射眼睛後由一般攝影機擷取眼睛的反射影像，但一般光線較無法突顯出眼睛特徵，所以對後續的處理會造成一定程度的困難。對此本論文提出三點測圓法來追蹤角膜緣，也就是虹膜邊界，從一般光線反射的眼部影像中搜尋出虹膜邊界上的三個特徵點，並以此三點為頂點之三角形的外接圓預估出虹膜邊界及中心點。此外若有一眼無法成功估測出虹膜邊界，則可以成功估測出之眼睛的結果為基礎，用映射補償的方式去推算出另一眼的結果。三點測圓法雖然在步驟上有些複雜，但幾乎都是運算量較低的簡單運算，另外映射補償的方式在計算上也不複雜，但成功映射出的機會很高。 實驗顯示處理每一張影像的平均時間為42毫秒。此外，虹膜邊界偵測成功的平均機率為99%以上，最低時也有96%的機率，而雙眼皆抓取到的平均機率也在98%以上。映射到螢幕上的落點誤差在水平方向的平均誤差為4.5%，垂直方向的平均誤差為3.5%，可以證明此系統在抓取異色邊界地帶為有效且可靠的。
並列摘要
Eye tracking has many applications. For example, some smart phones can detect whether human eyes are looking at the screen of the smart phone; if not, the video playback will be paused automatically. In addition, some computer makers install an infrared LED on Tablet PCs to detect eye gaze positions and record the position data. Advertising companies can use these data to do ad-benefit analysis. Without installing the infrared LED, those devices can only detect whether a user is looking at the screen, and the eye gaze detection is not possible. Installing an infrared LED may enable the estimation of eye gaze positions, but the LED will make the user’s eyes feel dry and even got hurt. There are many ways to do eye tracking. They can be divided in two categories, namely contact and non-contact. For example, the EOG (electro-oculogram) method and the search coil method belong to the contact category, while optical-type eye tracking and infra-red oculography belong to the non-contact category. Unfortunately, these methods may directly or indirectly cause the physical harm for eyes. Among all methods the search coil method is the most invasive, where a local anesthetic on the eye part is required. For the non-contact category, most methods use infrared LEDs to make the eye pupil more obvious, but it can also cause dry eyes and uncomfortableness. To avoid the damage to the eyes when we use invasive devices or devices with infrared LED for eye tracking, we can use a fluorescent light to illuminate the eyes naturally and capture the eye reflection image by an ordinary camera. However, the fluorescent light normally can not highlight the eye feature, presenting some difficulty in subsequent processing. This thesis proposes a method called three-point circle detection to track the limbus or iris boundary. In the proposed method, three iris feature points on the iris boundary are found from the fluorescent light reflection of the eyes and the boundary as well as the center of the iris are estimated as a circumscribed circle determined by the three points. In addition, when the estimation fails for one eye, its estimation result can be derived from the successful estimation result of the other eye based on an idea of mapping compensation. The three-point circle detection procedure may be somewhat complex, but only simple calculations with low computational cost are involved. The approach based on mapping compensation is also computationally tractable and it often produces good mapping results. Experimental results show that the average time to process each image is 42 milliseconds. In addition, the average successful detection rate of iris boundary is more than 99%, with 96% the lowest. The average successful detection rate for both eyes is more than 98%. The average position error of gaze point on a screen in the horizontal direction is 4.5%, while it is 3.5% in the vertical direction, which can show that the proposed methods are effective and reliable in detecting the limbus area.