Recognition of eye movement based on bioelectrical signals using neural networks

Abstract

The object of research is the process of generating and recording electrical signals caused by eye movements; the subject of research is the method of real-time recognition of eye movements based on these signals. It is implemented on the open VitalCore platform and uses a convolutional neural network (CNN) for real-time movement classification. One of the most problematic aspects is ensuring high accuracy with low power consumption and limited computing resources, as well as reducing the impact of noise and delay during signal processing. This is of particular importance when using the system in wearable devices and in real-world environments where signal quality may be unstable.

The study uses digital signal processing methods, in particular, filtering by the Savitsky-Goley algorithm, as well as architectural solutions in the field of neural networks: the use of a five-channel CNN with ordinary and transposed convolutional layers, Flatten and softmax. The use of frequent sliding windows (every 8 ms) is proposed, which increases accuracy and reduces latency.

The result is obtained: the recognition accuracy reaches 85% with a time window of 625–833 ms and a latency of about 40 ms, which provides the ability to detect up to five movements per second. This is due to the combination of an energy-efficient sensor with an optimized CNN architecture, which provides noise immunity and fast classification in real time.

Thus, the method allows to achieve stable and reliable results while maintaining low power consumption. Compared with known analogues, it is distinguished by openness, scalability, reproducibility and the ability to work on peripheral devices without high-performance computing resources. The development can be integrated into wearable devices and used in brain – computer interfaces, VR/AR, assistive technologies and medical research, which emphasizes its practical value.