Pages Navigation Menu
Submit scientific paper, scientific publications, International Research Journal | Meždunarodnyj naučno-issledovatel’skij žurnal

Выпуск № 06 (120) Июнь 2022

ISSN 2227-6017 (ONLINE), ISSN 2303-9868 (PRINT), DOI: 10.18454/IRJ.2227-6017
ЭЛ № ФС 77 - 80772, 16+

PERCEPTUAL INTEGRAL PATTERNS AS A METHOD FOR RAPID PUPIL TRACKING BASED ON ANN-MLP PERCEPTRON RESPONSE AND A MODIFIED NUMERICAL METHOD OF DIFFERENTIAL EVOLUTION

Posted in 2022, ENGINEERING, TECHNICAL SCIENCE, Выпуск № 06 (120) Июнь 2022 | 0 comments

The article offers a fast and accurate model of registration and tracking of pupils «RAPID-AN» in the task of video-oculography for devices with a limited computing resource (weak personal computers, embedded systems based on ARM architecture), to reduce the technology cost for individual use by disabled people and by health facilities. The model was based on the idea of the pupil registration process in the frame (finding the center coordinates and the region size most fully capturing the pupil) as a global, multifaceted optimization task. The target function is formalized as the meaning of ANN-MLP response perceptron over integral patterns, in a vivid sense similar to the desired object structure (pupil-iris) – the concept of perceptual integral patterns. In order to improve the convergence to extremum, a modification of differential evolution method is proposed, based on the process of forming the genetic isolation of solutions population in the vicinity of all local and global extremes of the target function, with subsequent growth of the most adapted insulation (next to the global extremum) and the degeneration of others, according to the Verhulst-Pierre differential equation. This behavior makes the search method less «greedy» and makes it possible to correctly extract a pupil from the full frame.

The model solves the problem of pupils registration with an error of not more than 1 – 2% on test datasets and a high speed of ~ 8.35 ms. in tracking mode on the architecture of BCM2837 ARM Cortex-A53 1200 MHz (embedded systems).

The comparative analysis was conducted, and performance of the model in comparison with some existing detectors was shown: Tiny Yolo v3, Viola-Jones (Haar-Cascade) method, etc.

Read More