Multi-Directional Pixel Array Computation for Efficient Pupil Localization in Gaze Estimation

Accurate pupil localization is a crucial requirement in gaze estimation, which can be employed to enhance the performance of human-computer interactions. Due to the complexity of practical scenarios, as well as the real-time and accuracy requirements of calculation, accurate pupil localization with robustness and real-time is still a challenging pattern recognition task. This study proposes a new pupil localization method based on Multi-directional Pixel Array Computation (MPAC), together with several optimization strategies including rotation transformation and pixel-array rotation similarity index. These strategies can further improve the localization accuracy by regulating pixel array distribution and fine-tuning the initial localization, respectively. For the application in gaze estimation, the proposed pupil localization method is encapsulated as a light-weight module and integrated into the gaze estimation model. To assess the accuracy and real-time performance of the proposed pupil localization method and compare it with the currently available methods, experiments on the BioID benchmark dataset were implemented. Furthermore, the training and test experiments of gaze estimation models were carried out to verify the contribution of the pupil localization module in improving the performance of gaze estimation. Experimental results show that the proposed method for robust and efficient pupil localization can be employed to achieve comparable performance with state-of-the-art learning-free methods. It has less computational complexity and achieves nearly real-time pupil localization with around 60 frames per second on the BioID benchmark dataset. Moreover, the proposed pupil localization method can be encapsulated as a callable module to improve the test accuracy of the gaze estimation models.