Pupil-Tracking-Project

This project aims to track the pupil boundary in real-time.

A quick demo is found here.

https://youtu.be/3MKtuQSaXd4

Setups

• Software: Windows 8, OpenCV 3.10 and Pyhton 2.7
• Hardware: IR camera, resolution 640x480.

Pipeline

Pre-process to crop eye region

• Read video streaming from camera frame by frame
• Grayscale
• Smooth image with median blur filter. For this application, median blur works better than gaussian blur.
• Apply Haar Cascades frontal face detector[1]
• Crop eye region with geometry assumption that eye location is relatively fixed to face.

Apply algorithms to fit pupil boundary with ellipse (Demo for one pupil only)

• Compute sobel directions of gradient
• Canny edge
• Find contours on the canny edges. Discard contours with large arc perimeter.
• Compute gradient entropy (refer to Cihan Topal's publication[2]). Select maximum entropy among all contour candidates.
• Fit ellipse for the contour

Reference

• [1] Apply Haar Cascades frontal face detector Download xml file from https://github.com/opencv/opencv/blob/master/data/haarcascades/haarcascade_frontalface_default.xml Note the license using the Cascades.
• [2] C.Topal and C. Akinlar, An Adaptive Algorithm for Precise Pupil Boundary Detection Using the Entropy of Contour Gradients.2013. Elsevier preprint.