Robust Webcam-Based Gaze Tracking Using Double-Eye Convolutional Networks with Kalman Filtering and Clustered Calibration

Webcam-based gaze tracking has emerged as a cost-effective alternative to infrared eye-tracking systems; however, its robustness under real-world conditions remains limited, especially when monocular (single-eye) models are employed. This study offers a systematic comparison between single-eye (SET) and double-eye (DET) convolutional neural network architectures for appearance-based gaze estimation utilizing standard webcams. A lightweight CNN processes normalized eye-region images to predict gaze coordinates on the screen and is evaluated through two complementary protocols: (i) a static 9-point calibration and (ii) a dynamic trajectory-following task (“blue-ball”) that traverses screen edges and corners. To enhance generalization, the dataset incorporates controlled data augmentation—pose variation, illumination changes, blur/noise, and partial occlusions—and employs a participant-stratified experimental design. Robustness and temporal consistency are further improved by integrating a Kalman filter for trajectory smoothing and a DBSCAN-based clustering calibration stage that suppresses outliers and stabilizes gaze estimates. Performance metrics include pixel error, angular error, and trajectory stability. Under identical training and evaluation conditions, the DET model consistently surpasses the SET model, attaining lower spatial error and smoother trajectories, particularly under asymmetric lighting and partial occlusions. Utilizing only a consumer-grade webcam and a lightweight CNN, the proposed methodology delivers real-time performance without infrared sensors or proprietary licenses, thereby providing an accessible, reproducible, and robust solution for gaze estimation in human–computer interaction applications. Smart citations: https://scite.ai/reports/10.61467/2007.1558.2026.v17i2.1304Dimensions.Open Alex.