Hybrid iris center localization method using cascaded regression, weighted averaging, and weighted snakuscule

Abstract

Iris center estimation is widely used in various computer vision applications, such as driving monitoring and eye tracking. However, accurately locating iris centers in low-resolution images remains a significant challenge. We propose a robust, accurate, and real-time iris center localization method based on cascaded regression, weighted averaging, and weighted snakuscule. In the proposed scheme, a powerful cascaded regressor is trained to detect the eye contours and iris centers, which is further refined by the inverse-intensity weighted averaging method. Further, an improved weighted snakuscule is proposed to fine-tune the detected iris centers. The performance of the proposed method is tested on publicly available databases, namely BioID, GI4E, and Talking Face Video. Accuracies of 96.58%, 98.30%, and 96.12%, respectively, are achieved at a normalized error <0.05. Compared with the state-of-the-art methods, the proposed scheme increases the overall accuracy by 3.72% at a normalized error <0.025 and achieves the highest accuracy on the BioID and Talking Face Video databases. The total execution speed is 33 fps. The superior performance of the proposed method proves its usefulness for real-time application with improved robustness and accuracy.