by wxz » 19 Aug 2014, 23:44
It won't hurt to try out Kalman filtering, using openCV for example. And it may provide some improvements. But I'm afraid it won't fix the key issues causing inaccurate eye tracking. For example, change of head position, distance, head orientation etc.
Usually, the mapping from the eye position on the infrared image to the gaze position on the screen is done with non-linear, approximated polynomials (not an exact dynamic model). Simply filtering the numbers generated from such mapping cannot correct the systematic error.