Adaptive EOG signal processing on DSP

Type: Master thesis
Students: Roman Suter
Advisor: Andreas Bulling, Clemens Lombriser
Project: Eye Movements and Visual Perception

In this thesis, an integrated, wearable EOG recording device was equipped with a framework to process electrooculogram (EOG) signals directly on the device. EOG signals are sampled and intended to be processed in real-time. Therefore, the framework design considered the constraints of computational power of a DSP, the few memory access for data processing and the real-time feasibility of the included filters.

In a first step, the framework was made remotely controllable over the Bluetooth link by redefining the configuration packet of the device. Single filters can be turned on and off without a recompilation and filter parameters can be set to calibrate the current session. Additionally, some filters are available to be set in a desired order.
An eye movement event encoding was introduced to merge the vertical and horizontal EOG channel or to discard wrong event postings due to artefacts. The former enables the detection of cross direction eye movements and the latter reduces the error-event rate significantly.

An effort was done to bring the eye tracker together with body movements. Step shocks cause strong artefacts in the EOG signal and complicate the detection of characteristical eye movements. The artefacts were analysed and an approach with an adaptive median filter was proposed to suppress the undesired contamination of the signal. Acceleration signals were used to classify the current movement state of the user and to find the optimized median window size during walking activity. An experiment proved the improvements of the EOG signals due to the adaptive median filter.