Hybrid LASSO and Neural Network Estimator for Gaze Estimation

Gaze estimation has wide applications in drowsiness detection, security, and biomedical domains. The challenges in estimating the gaze angle include varying light conditions and subtle movements of the gaze. The Convolutional Neural Network (CNN) has recently been suggested as a potential method for gaze estimation. In this present work, we have proposed a gaze estimator combining a neural network and Least Absolute Shrinkage and Selection Operator (LASSO). The features considered are both eye and head features. The combined estimator neural network-LASSO (NN-LASSO) outperforms the individual performance of neural network and LASSO estimator. The results are validated using MPH Gaze dataset and it has been shown that the proposed NN-LASSO estimator outperforms CNN in mean error sense.