Embedded CPU-GPU pupil tracking

We explore camera-based pupil tracking using high-level programming in computing platforms with end-user discrete and integrated central processing units (CPUs) and graphics processing units (GPUs), seeking low calculation latencies previously achieved with specialized hardware and programming (Kowalski et al., [Biomed. Opt. Express12, 6496 (2021)10.1364/BOE.433766]. Various desktop and embedded computers were tested, some with two operating systems, using the traditional sequential pupil tracking paradigm, in which the processing of the camera image only starts after it is fully downloaded to the computer. The pupil tracking was demonstrated using two Scheimpflug optical setups, telecentric in both image and object spaces, with different optical magnifications and nominal diffraction-limited performance over an ∼18 mm full field of view illuminated with 940 nm light. Eye images from subjects with different iris and skin pigmentation captured at this wavelength suggest that the proposed pupil tracking does not suffer from ethnic bias. The optical axis of the setups is tilted at 45° to facilitate integration with other instruments without the need for beam splitting. Tracking with ∼0.9-4.4 µm precision and safe light levels was demonstrated using two complementary metal-oxide-semiconductor cameras with global shutter, operating at 438 and 1,045 fps with an ∼500 × 420 pixel region of interest (ROI), and at 633 and 1,897 fps with ∼315 × 280 pixel ROI. For these image sizes, the desktop computers achieved calculation times as low as 0.5 ms, while low-cost embedded computers delivered calculation times in the 0.8-1.3 ms range.