Badr OUANNAS: Dualité fixation-poursuite en suivi des mouvements oculaires : Analyse des perceptions visuelles en temps-réel sur objets mobiles

De Ensiwiki
Aller à : navigation, rechercher


Cornues.png
Titre du projet Dualité fixation-poursuite en suivi des mouvements oculaires : Analyse des perceptions visuelles en temps-réel sur objets mobiles
Cadre IRL

Labo LIG

Encadrants Francis JAMBON


An eye-tracker in simple terms is a device that traces the movements of the eye on a display in order to analyze the behavior of the user's perception responding to certain stimuli. The applications of eye-Tracking are varied from applications in sport such as: analyzing the strategies of expert chess player vs beginners, learning simulations like: driving simulations and even psychology.

Eye movement can be classified to three main classes: Fixation, Pursuit and Saccade. Fixation is as its name suggests is when the gaze is still and focused on the stimuli, pursuit on the other hand is when the gaze follows the movement of an object, like when following the movement of a bird in the sky, the movement is not too fast that it would make the gaze jump from a position to another. But when it is the case of high movement or when the gaze jumps from one position to another quickly we have then a saccade movement, like when reading through a book the gaze jumps from line to line.






The goal of the Project

The goal is to identify algorithms for eye movement detection in real time. There exist many algorithms to classify eye-movements, but most of them can be categorized in one of two approaches: Probabilistic approach and Threshold-based Approach. The threshold method use the velocity threshold to classify the eye movement: a fixation would have a very low velocity threshold, as for the saccade its velocity will be around 100◦/s, thus the pursuit movement can be classified by using both thresholds. In addition to using a velocity threshold: many algorithms use the duration of the eye movement for classification (such as I-VDT). With the probabilistic approach we use techniques such as Bayesian decision theory to identify the movement depending on the data provided by the eye-tracker.

Let S = {(xi, yi, ti)/0≤i≤N} be a set of N+ 1 ordered tuples. Where(xi,yi) represents the position of the gaze on the display and ti represents a time stamp. The problem is to classify each period between two subsequent tuples to either fixation or pursuit. And for that we will be using two approaches: I-BDT (Bayesian Decision Theory Identification) and a threshold-based approach.    

The Approaches

The Bayesian Approach

Using Bayesian Decision Theory (or BDT for short) means to find the probability of Event knowing the Data: P(Event/Data), where Event is either Fixation or Pursuit. BDT allows us to find this probability using the following Bayes rule:

p(Event/Data) = p(Data/Event)×P(Event)/P(Data)

Thus to find the posterior probability P(Event/Data) we need to define each term in the formula: the prior(Event), the Likelihood P(Data/Event)and the scaling probability P(Data). The calculations are based on a model well defined in the research paper "Bayesian Identification of Fixations, Saccades, and Smooth Pursuit". Upon finding the posterior we can proceed to the classification.

The Threshold Approach

Using the angular threshold for both fixation and saccade we can calculate the threshold of the velocity on the display, and by using this velocity and compare it to the current velocity calculated using the data retrieved from the eye-tracker, we can classify at every iteration the movement of the eye; If it is below the threshold of fixation that means that the movement is certainly a fixation, otherwise it is probably a pursuit.

The Experimentation

Fast movement of the stimuli in the experimentation

The experimentation involved 7 people (4 men and 3 women), it required them to follow the movement of a stimuli on a black background. the stimuli used is a white dot with a cyan colored bull's eye in the center. We covered a range of 3 main movements for this stimuli in two different velocities (slow and fast): horizontal, vertical and the diagonal. For each type of movement the stimuli would simulate a pursuit followed by a fixation then followed by a pursuit until it gets to the border of the display, then it would repeat the same thing in the reverse trajectory.

We evaluated the algorithms using four metrics: Sensitivity (TP/(TP+FN)) , precision (TP/(TP+FP)), specificity (TN/(TN+FP)) and accuracy ((TP+TN)/(TP+TN+FP+FN)), Where TP, FP, TN and FN stand for True Positive (correctly classified pursuit), False Positive (incorrectly classified pursuit), True Negative (correctly classified fixation) and False Negative (incorrectly classified fixation).

The code for the experimentation and the results for the different approaches used can be viewed in the rapport below.
The eye Tracker used is a Tobii 4C (sampling frequency of 90 Hz) with Tobii Pro SDK upgrade and a Dell 2416D monitor (refresh rate of 60 Hz

Links

  • Rapport : PDF
  • Thiago Santini, Wolfgang Fuhl, Enkelejda Kasneci, Thomas Kubler: Bayesian Identification of Fixations, Saccades, and Smooth Pursuits ([1])
  • The SDK for Tobii Pro:[2]