Gazing estimation and correction from elliptical features of one iris

The accuracy of eye gaze estimation by image information is affected by several objective factors, including the image resolution, anatomical structure of eye, posture change, etc. Especially, the irregular movements of head and eye are the main problem and key technology being researched. We describe an effective way of estimating the eye gazing from the elliptical features of one iris under the conditions without auxiliary source, head fixing equipment or multiple-camera. Firstly, we give the preliminary estimations of the gazing direction and then obtain the vectors describing translation and rotation of eyeball movement using central projection on the cross section passing through the line-of-sight, which avoids the complex computations involved in known methods. We also disambiguate the solution on the basis of the experimental findings. Secondly, the error correction is carried on the BP neural network trained by a sample collection of the translation and rotation vectors. In our simulations, we achieve an accuracy of 0.8° on the test images which are different from the training images. The result is found to be better than that of the existing non-intrusive method with single-camera. The performance of the algorithm proves that the proposed method has excellent generalized abilities.