Method and device for carrying out eye tracking
A method is disclosed for determining an eye position and/or eye gesture, in particular for use in controlling a function of an application system, in particular a wearable. The method includes (i) recording a profile of at least one EEG signal over time by way of a contact electrode in order to obtain a signal time series of the at least one EEG signal, (ii) providing input data sets from a signal block which arises from the signal time series, for successive time windows, and (iii) evaluating the input data sets in a data-based eye movement model in order to obtain the eye position or eye gesture, the eye movement model being formed using a neural network with one or more recurrent layers, in particular a respective LSTM layer or GRU layer or a transformer layer, and one or more neuron layers.
1 . A method for determining an eye position and/or eye gesture for use in controlling a function of a wearable, the method comprising:
acquiring a profile of at least one EEG signal over time by way of a contact electrode to obtain a signal time series of the at least one EEG signal;
determining a plurality of input data sets by (i) dividing the signal time series into a plurality of successive time windows, (ii) extracting respective signal characteristics from each of the plurality of successive time windows, and (iii) forming each respective input data set in the plurality of data sets to include a respective time window from the plurality of successive time windows and the respective signal characteristics extracted from the respective time window; and
determining a respective eye position and/or a respective eye gesture for each respective time window in the plurality of successive time windows by evaluating each respective input data set in the plurality of data sets using a data-based eye movement model, the data-based eye movement model being a neural network having one or more recurrent layers and one or more neuron layers and being configured to output the eye position and/or the eye gesture.
2 . The method according to claim 1 , wherein:
the signal time series is processed to remove interference,
the interference includes a jitter, a potential trend, and/or a noise component and is removed or reduced, and
the input data sets are determined from provided from the processed signal time series.
3 . The method according to claim 1 , wherein:
the eye position is provided in a two-dimensional eye coordinate system, and
an x-value indicates a horizontal eye position and a y-value indicates a vertical eye position.
4 . A device for carrying out the method according to claim 1 .
5 . An application system, comprising:
one or more attached contact electrodes configured to acquire a respective EEG signal; and
the device according to claim 4 .
6 . The method according to claim 1 , wherein the one or more recurrent layers includes a respective LSTM layer or GRU layer or a transformer layer.
7 . The method according to claim 2 , wherein the interference is removed or reduced by at least one of linear regression, trend elimination, and noise elimination.
8 . The application system according to claim 5 , wherein the application system includes a head-mounted device.
9 . The application system according to claim 8 , wherein:
the head-mounted device includes smart glasses,
the smart glasses include (i) one or more contact electrodes, and (ii) a side piece and/or a nose pad, and
the one or more contact electrodes are disposed on the side piece and/or on the nose pad.
10 . A method for training a data-based eye movement model to determine an eye position and/or eye gesture for use in controlling a function of a wearable, wherein the eye movement model is a neural network comprising one or more recurrent layers and one or more neuron layers, comprising:
creating training data sets, wherein the training data sets are respectively created by:
acquiring a profile of at least one EEG signal over time by way of at least one contact electrode to obtain a signal time series of the at least one EEG signal;
determining a plurality of input data sets by (i) dividing the signal time series into, a plurality of successive time windows, (ii) extracting respective signal characteristics from each of the plurality of successive time windows, and iii) forming each respective input data set in the plurality of data sets to include a respective time window from the plurality of successive time windows and the respective signal characteristics extracted from the respective time window,
determining, for each respective time window of the plurality of successive time windows, a respective eye position and/or a respective eye gesture during the respective time window; and
forming the training data sets from the respective input data set and the respective eye position and/or the respective eye gesture, of each respective time window of the plurality of successive time windows; and
training the eye movement model, with the training data sets, to determine an eye position and/or eye gesture.
11 . The method according to claim 10 , wherein the one or more recurrent layers includes a respective LSTM layer or GRU layer or a transformer layer.