A High-Frequency SSVEP BCI System for Natural Interaction
This project comes from an actual research project Contextual Computing Group. The core of the work involves designing and optimizing an 8-channel, 500Hz dry-electrode BCI system, utilizing a steady-state visual evoked potential (SSVEP) paradigm for real-time multi-class classification and control.
1. Platform Building
The experimental system was built on an 8-channel OpenBCI board, with data transmitted via WebSocket. To address the insufficient temporal resolution of the original OpenBCI system, the timestamp recording method was redesigned using an Arduino. Additionally, after data acquisition on the server, spline interpolation was applied for resampling to ensure uniform data spacing.
2. Algorithm Optimazition
We attempted to establish an event classification mechanism using EEG band-specific energy, EEG signal-to-noise ratio (SNR), and low-frequency energy. Through small-scale data collection and offline experiments, we determined the trigger thresholds, achieving a seated posture recognition accuracy of 85% with a latency under 5 seconds, and a standing posture recognition accuracy of 75% with a latency under 10 seconds.
3. Small-scale Data Acquisition
We conducted a small-scale data acquisition study with five participants to optimize the classification algorithm, testing both seated and standing postures. We found that the duration of data collection, reflecting participants' levels of concentration, influenced the distribution of trigger thresholds.
Main Contributions:
1. Developed a Python-based WebSocket Server/Client system integrating Arduino-controlled visual stimulation with real-time EEG acquisition, classification, and feedback via OpenBCI, achieving low-latency predictive output within 4 seconds;
2. Applied filtering and resampling to real-time 8-channel SSVEP signals, constructing 4-second FFT windows with 1/8 rolling step size, supporting a minimum 5Hz classification interval. Achieved real-time classification accuracies of 75% for standing posture with an average latency of 10s, and 85% for seated posture with an average latency of 5s.
