Final Project for Computer Vision Eye Feature Detection Rui Liu PhD student M.E. CONTENT ‣ 1 Introduction ‣ ‣ ‣ 1.1 Motivation 1.2 Goals 1.3 Experiment Devices and Environment ‣ 2 Methods & Algorithms ‣ 3 Experiment Results ‣ 4 Result Analysis ‣ ‣ ‣ 4.1 Assessment 4.2 Small Problems and Its Analysis 4.3 Moving Forward 1 Introduction 1.1 Motivation for Eye feature detection (1) What could eye feature indicate? reflecting psychological state He sees the truth. It’s written all over our faces. Indicating human intention Maybe want to leave Maybe want to drink (2) Motivation: detecting eye features for intuitive human-robot interaction 1.2 Goals Pupil detection pupil localization, pupil diameter/ mean area and its standard deviation in a certain time. Blink detection blinking status detection, blinking times and blinking rate in a certain time. 1.3 Experiment Devices and Environment (1) Devices HD Camera Pic Captured Head Mount (2) Environment Common lab (400 lux) 2 Methods & Algorithms 1.1 pupil detection (1)Method: a. Circle(d≈44) b. Dark(Area≈1500) c. Concentric d. Region Pupil (2) Algorithm: a. Circle detection: Two-stage Circular Hough Transform (‘imfindcircle’ ) b. Dark region detection: ‘regionprops’ Concentric d= 42 ‣ (3) Parameter Calculation pupil center (x, y) circle detection Pupil Diameter Mean Area A = 𝑡2 𝜋𝑑2 𝑡 𝑡1 4 d /(𝑛𝑡2 − 𝑛𝑡1 ) Standard Deviation of Pupil Diameter 𝜎 = 𝑡2 𝑡1(𝑑 𝑛𝑡 −𝑑)2 𝑛𝑡2 −𝑛𝑡1 1.2 blink detection (1) Method: pupil disappearing time >𝒕𝟎 Blinking Blinking Open ‣ (2) Parameter Calculation Blinking Times N Blinking Rate Rate=N/(𝑡2 -𝑡1 ) 3 Experiment Results 3.1 results in different situations a. Beginning b. Blinking c. Open d. Near Blinking 3.2 Result Video 3.3 Statistics Results blink rate Blinking Status 1.5 1 -- Blinking 0 -- Open Eye Status 1 0.5 0 -0.5 0 5 10 Time(s) a. Eye status 15 20 blink rate(Times/sec) 1.5 1 0.5 0 -0.5 0 5 10 Time(s) 15 b. Blinking Rate (n/sec) 20 Pupil Diameter Mean Pupil Area Mean Pupil Area(pixel 2) 40 20 1600 1500 1400 1300 0 0 0 5 10 Time(s) c. Pupil Diameter 15 5 10 Time(s) 20 Standard Deviation of Pupil Size Pupil Diameter(pixel) 60 3 2 1 0 5 10 Time(s) 15 20 d. Mean Pupil Area Standard Deviation of Pupil Size 0 15 20 e. Standard Deviation of Pupil Size 4. Result Analysis 4.1 Assessments Successful Rate for Pupil detection 99% Successful Rate for Blink detection 100% Successful Rate for Pupil Contour detection 90% 4.2 small problems and its analysis Open (a) A few pupil contours could not be detected accurately when eye is near blinking/open N R Because Actual contour shape is ellipse which is detected as 2 circles — A (b) Pupil in some frames could not be detected near blinking/open. Because Dark region detection : Failed Light intensity is too low near blinking/open. 4.3 Moving Forward In the future, this project could be optimized in next aspects: (1) The circle detection could be revised as circle/ellipse detection (2) The sensitivities and thresholds of these algorithms could be adjusted adaptively Q&A Thank You ! Rui Liu