Improvement Of An Untrained Brain-computer Interface System Combined With Target Recognition

In the current commonly used Steady State Visual Evoked Potential (SSVEP) paradigm, the stimuli are mostly white flashing blocks superimposed on a black background, which is monotonous and easy to cause subject fatigue with prolonged flashing stimuli. The stimulus paradigm is mostly divorced from the actual control environment, and lacks a direct connection with the control task. The mainstream classification algorithms usually analyze the data with a fixed window length, which is lack of generalizability to different subjects, and the classification performance index needs to be further improved. In this study, the SSVEP stimulus paradigm was improved by combining the YOLOv5 algorithm, which changed from the traditional black background to the actual control environment. It superimposed SSVEP stimulus blocks of different frequencies at each recognized target location. The stimulus paradigm was not stripped from the control scene, and the Filter Bank Criterion Correlation Analysis (FBCCA) algorithm was chosen to analyze it. The FBCCA algorithm was further improved by using a dynamic window strategy, which automatically adjusts the window length of each experiment according to the characteristics of each subject. This improves the versatility of the algorithm and increases the recognition accuracy and Information Transfer Rate (ITR). After the improvement, the offline experimental data were analyzed. The improved algorithm achieved an average accuracy of $87.08 \%$, which was $17.29 \%$ higher than the original algorithm. Additionally, the average ITR was 74.28 bits/min, which was $36.51~\mathrm{bits} / \mathrm{min}$ higher than the original algorithm.