Comparative Analysis of Filtering Techniques in Eye Landmark Tracking: Kalman, Savitzky-Golay, and Gaussian

This study benchmarks Kalman, Savitzky-Golay, and Gaussian filters to improve eye landmark detection crucial for online proctoring systems. Using eye-tracking video data, we evaluated filters on Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), computation time, and jitter reduction. The Kalman filter emerged superior with an MSE of 0.49, MAE of 0.48, RMSE of 0.70, and MAPE of 15.56%, alongside a 2.39% reduction in standard deviation, 4.73% in variance, and the quickest computation time of 0.02 milliseconds. The Gaussian filter reported an MSE of 0.49, MAE of 0.45, RMSE of 0.70, and MAPE of 14.48%, with modest jitter reductions of 0.96% in standard deviation and 1.90% in variance. In contrast, the Savitzky-Golay filter, with an MSE and MAE of 0.70, increased jitter, shown by a -1.50% change in standard deviation and -3.03% in variance, and took the longest computation time at 1.44 milliseconds. The results solidify the Kalman filter’s suitability for real-time, precision-demanding eye-tracking, while the Gaussian filter serves well for scenarios with relaxed accuracy needs. The Savitzky-Golay filter’s applicability seems limited due to its propensity to amplify noise under the conditions tested.