Classification of Visual Perception EEG Signals for a 2D Platformer Game

This research focuses on classifying visual perception brain signals using an OpenBCI headset to capture the EEG signals from the participants. The headset comprises 16 re-configurable dry channels, which poses two challenges: the amount of data and high noise levels. The primary objective is to classify the intent of player movements in a 2D platformer game with a convolutional neural network. Since every player has a distinct EEG footprint for similar intents, we explore the process of training a global model suitable for many players. During the experimental setup, participants were instructed to navigate the player character within the game by solely relying on visual perception, simulating a realistic gaming environment. By leveraging existing EEG models such as EEGNet and ShallowConvNet, along with a newly designed 1D CNN model (Conv1D), the research compares the accuracy of these models on individual participant datasets and pooled datasets. The results reveal that the Conv1D model outperforms the other models, achieving high accuracy (100%) for all individual participants. Moreover, when the models are trained on pooled data, the Conv1D model retains its high accuracy, while the accuracy of the other models significantly decreases. However, when the model was tested on datasets from participants not included in the training phase, suboptimal results were observed, with a maximum accuracy of 41%.