Robust real-time retinal tracking for ophthalmic applications

This work presents a software-based approach for robust feature tracking of funduscopy images. In our specific application, it is utilized in a scannerless linear optical coherence tomography system, where both natural and directed eye movements are used to scan the retina. The approach developed includes a fast, contrast-enhancing video preprocessing step. Additional filtering highlights edges and details, enhancing the visibility of blood vessels and the optic disc. Reflections are removed, both those arising from the funduscopy setup itself and non-stationary reflections caused by the cornea or an intraocular lens. A state-of-the-art feature detector and descriptor is used to identify and characterize distinctive image regions. Subsequent feature matching and filtering include additional criteria to enhance robustness against outliers and false detections. From these final matches, homographies are calculated, allowing the derivation of relative movements and absolute positions. The results demonstrate real-time processing with high detection rates and minimal misdetections. This performance is maintained even in the presence of poor contrast and non-stationary reflections in the original video stream. While the tracking is optimized for our application, it is also applicable to other domains, such as optimizing the alignment of retinal images, generating wide-field panorama images from individual frames, or characterizing eye movements.