Publikation
Local Optical Flow for Eye Movement Event Detection in Head-Mounted Setups
Abdulrahman Mohamed Selim; Omair Shahzad Bhatti; Lukas Wilde; Minh Khue Pham; Michael Barz; Daniel Sonntag
In: Proceedings of the 28th International Conference On Multimodal Interaction. ACM International Conference on Multimodal Interaction (ICMI-2026), located at ICMI-2026, October 5-9, Napoli, Italy, Pages 290-298, ISBN 9798400723186, Association for Computing Machinery, New York, NY, USA, 10/2026.
Zusammenfassung
Detecting eye movement events is essential for analysing visual behaviour and enabling gaze-based interaction. However, many algorithms are designed for stationary eye trackers and struggle in dynamic head-mounted settings, particularly for smooth pursuit detection. In this study, we extend a recent multi-step framework for head-mounted eye movement event detection by replacing globally averaged scene motion with gaze-centred local optical flow extracted using the NeuFlow v2 optical flow estimator. We evaluate this idea in two ways: with a threshold-based approach, and with a unified Random Forest classifier that uses local-flow features to capture object-relative motion. Using the Gaze-in-Wild dataset, and compared with established baselines from the literature, the proposed classifier improves smooth pursuit event-level F1 from (27.2\%) (the strongest baseline) to (44.9\%), and achieves the highest event-level weighted average (F_1 = 77.7\%), while remaining competitive for fixation and saccade detection. We also demonstrated real-time feasibility by integrating our method into a signal processing pipeline with ≈ 40 ms end-to-end latency at 30 FPS on a desktop GPU setup. Despite relying on a single dataset, we show that local optical flow improves head-mounted pursuit detection.