Head Tracking in Kabinenumgebung

Type: Semester Thesis
Student: Claudia Lorenz
Advisor: Johannes Schumm, Cornelia Setz
Project: SEAT

To provide a high comfort on long-distance flights context awareness is a very important topic. One possibility to achieve context awareness, is an analysis of the passengers' line of sight with continuous video recording of its head. Therefore an algorithm was developed as follows: First a head detection is performed, after that the eyes will be extracted and according to this the line of sight is detected.

The head recognition is based on the so-called Skin Color Matching that performs a comparison between the picture and an intersection histogram. Because various people have heads of different sizes, the comparison is done with changing sizes of cutouts. A problem, which can occur, is that the detected region of the head is too small. This can be avoid by choosing the 10 most likely candidates and take out the largest one. After the detection of the head the cutout can be improved by using face characteristics. Since the eyes are placed in the upper half of the face, the cutout can be reduced on the bottom by 40%.

The eye detection is based upon the gradient computation of black and white values. Because eyes are round many gradients are cumulated in the middle of the eyes. One can use this fact to detect the eyes: The eyes are most likely located where many gradients intersect. But due to the fact that there are other points in the face where a lot of gradients cross, e.g. at the nose, it is necessary to define assumptions where the eyes are located and positioned relatively to each other. A pair of candidates will be detected as eyes if the defined conditions holds best and the gradients crossings are maximum.

The last part of the algorithm contains the line of sight detection. First of all a dark square, which represents the iris, is searched. After that the algorithm looks for bright pixels right, left and below the square. According to the found pixels the line of sight is classified. For instance if the bright pixels are located right and left from the square, the line of sight “ahead” is detected.