A Novel SSVEP Brain-Computer Interface System Based on Simultaneous Modulation of Luminance and Motion

Steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) have received significant attention owing to their high information transfer rate (ITR) and low training requirements. Previous SSVEP-based BCIs mostly adopt the stationary visual flickers where only a few studies have explored the effect of moving visual flickers on the SSVEP-BCI. In this study, a novel stimulus encoding method based on the simultaneous modulation of luminance and motion was proposed. We adopted the sampled sinusoidal stimulation method to encode the frequencies and phases of stimulus targets. In addition to luminance modulation, at the same time, visual flickers also moved horizontally towards right and left at different frequencies (i.e., 0, 0.2, 0.4, and 0.6 Hz) following a sinusoidal function. Accordingly, a nine-target SSVEP-BCI was built to evaluate the influence of motion modulation on the BCI performance. Filter bank canonical correlation analysis (FBCCA) approach was adopted to identify the stimulus targets. Offline experimental results of 17 subjects revealed that the system performance decreased with the increase of superimposed horizontal periodic motion frequency. Our online experimental results showed that the subjects achieved 85.00 ± 6.77 % and 83.15 ± 9.88 % accuracy for the superimposed horizontal periodic motion frequencies of 0 and 0.2 Hz, respectively. These results verified the feasibility of the proposed systems. In addition, the system with 0.2 Hz horizontal motion frequency provided the best visual experience for subjects. These results indicated that moving visual stimulus can provide an alternative option for SSVEP-BCIs. Furthermore, the proposed paradigm is expected to develop a more comfortable BCI system.


I. INTRODUCTION
B RAIN-COMPUTER interface (BCI) is a technology that uses recorded brain signals to enable the interaction between the user brain and a computer, to manipulate the external environment [1], [2], [3], [4]. With recent advancements in brain science and computer technology, the field of BCI has also been developing rapidly [5], [6]. Electroencephalogram (EEG) has become the most commonly used signal for BCI studies, due to its non-invasiveness, ease of use, and low cost of corresponding equipment, and has been widely used in medical treatment, to support daily activities or to provide entertainment [7], [8], [9]. Among EEG-based BCIs, steady-state visual evoked potential (SSVEP)-BCIs have gained a lot of attention and widely used for cognitive abilities evaluation or robotic arm control [10], [11] because of high information transfer rate (ITR) and low training requirements [12], [13].
Currently, SSVEP-BCIs are being explored primarily in two directions: (1) further expanding the number of encoded targets. For example, Chen et al. [14] realized a calibration-free 160-target SSVEP-BCI system by extending the method of multi-frequency sequential coding. More recently, Chen et al. [15] proposed a spectrally-dense joint frequency-phase modulation encoding method and utilized it to achieve a 120-target SSVEP-BCI. (2) improving the performance of target recognition algorithms. For example, the filter bank canonical correlation analysis (FBCCA) divides the SSVEPs into multiple sub-band components, to utilize the harmonic frequency components in addition to fundamental frequency components, which enhances the recognition accuracy of SSVEP [16]. In particular, a 40-target SSVEP-based BCI developed using the ensemble task-related component analysis (TRCA)-based method realized an ITR of 325.33 bits/min [17]. In a recent study, task-discriminant component analysis (TDCA) proved to be a superior alternative to ensemble TRCA [18]. In summary, the performance of SSVEP-BCIs has been significantly improved over the recent years, by increasing the number of encoded targets and improving the performance of decoding algorithms.
Nowadays, the most existing SSVEP-BCI studies adopt the traditional stationary visual stimulus design to elicit the SSVEPs. Meanwhile, only a few studies have investigated the moving visual stimulus to elicit the SSVEPs. For instance, Kanoga et al. [19] used a single moving visual flicker to study the influence of head movement on the system performance, under various conditions including different moving directions, moving speeds, and flickering frequencies.
The results indicated that the system performance deteriorates with the increasing head moving speed, and the system performance under horizontal moving condition surpasses that under vertical moving condition. These results have suggested that a horizontal movement-based method that has less influence on the system performance should be elected to build practical SSVEP-BCI systems under moving condition. In another work, Duan et al. [20] obtained the optimal parameters to build a 3 × 3 moving visual stimulus paradigm after investigating the system performance at four different moving speeds and different phase intervals in the vertical direction using squares as the visual stimuli. A comparison between the stationary and moving paradigms revealed that the system performance of moving paradigm is comparable to that of the stationary paradigm. Progressively, Punsawad and Wongsawat [21] designed a novel method which uses a flickering strip that flashes left and right to give the illusion of strip movement, instead of the traditional stationary flicker, to reduce the eye fatigue. The obtained results showed that the recognition accuracy of two classes (left or right) based on SSVEP is about 80 %. At the same time, a hybrid system was designed to control the wheelchair to move forward, backward, left and right, where the overall classification accuracy of the system was 85.62 %, demonstrating about 13% improvement compared to the traditional paradigm when the subjects were under fatigue [22]. However, only one flicker frequency was used for the two-class classification, and further investigation is needed for multi-classification systems with different flicker frequencies. Likewise, Zhang et al. [23] used 12 flicker frequencies (i.e., 6.2-16.6 Hz with an interval of 0.9 Hz) to encode the moving visual stimulus with the same moving speed but with a random moving direction, to develop a novel SSVEP-based BCI. The results showed the average recognition accuracy of 86.67 % for the random motion paradigm, which is less than the average recognition accuracy of 89.26 % for the stationary paradigm. However, this change in performance may be due to the overlap of stimulus targets during the motion. In general, the literature on SSVEP-BCIs based on moving visual stimulus is scarce, and there is less consensus among the few reported studies. But research in this area is meaningful. On the one hand, Motion is a fundamental visual attribute [24] and individuals who lack motion perception have been found to live in a very different world of frozen images, where simple tasks like filling a kettle or crossing the road take on alarming difficulties [25]. The paradigm based on simultaneous modulation of luminance and motion may be able to be used to assess whether a patient has motion perception. On the other hand, studying the effect of motion modulation on the SSVEP-BCI performance would certainly enhance our overall understanding of the SSVEP-BCI systems.
In this work, we introduced a novel stimulus encoding method based on the simultaneous modulation of luminance and motion. The sinusoidal sampling stimulus encoding method [26] was adopted to realize stimulus frequencies and phases, while luminance change was realized by the joint frequency-phase modulation (JFPM) method [27]. In addition to luminance modulation, at the same time, the visual flickers also moved towards right and left in a horizontal direction at different frequencies (0, 0.2, 0.4, and 0.6 Hz) following a sinusoidal function. Besides, filter bank canonical correlation analysis (FBCCA) approach was employed to detect SSVEPs [16]. Furthermore, offline experiments were implemented to optimize the BCI system parameters, whereas online experiments were carried out to validate the feasibility of the proposed novel stimulus encoding method. Meanwhile, the subjective comfort scores were also collected to further estimate the feasibility of the BCI paradigm.

II. METHODS AND MATERIALS
A. Experimental Environment 1) Subjects: Twenty participants (3 males and 17 females, with ages ranging from 21 to 30 years) joined this study. The overall study was set up with both offline and online experiments. Specifically, 17 participants (3 males and 14 females, age ranging from 23 to 30 years, with average age of 24 years) joined the offline experiment while 12 participants (3 males and 9 females, age ranging from 21 to 30 years, with average age of 25 years) participated in the online experiments. Moreover, 9 participants participated in both offline and online experiments. Besides, subjects were informed of the experimental task prior to the experiment, and they signed a consent form and obtained the payment after the experiment. The participants were physically and mentally healthy, with normal or corrected vision. During the experiments, subjects were asked to sit in a relaxed state in a chair which was 60 cm away from the screen. Approval of all ethical and experimental procedures and protocols of this study was granted by the Institutional Review Board of Tsinghua University.
2) Experiment Equipment and Data Acquisition: The Synamps2 system of Neuroscan was adopted to record the scalp EEG data at a sampling rate of 1000 Hz. Next, the collected EEG signals were band-pass filtered from 0.15 Hz to 200 Hz and trapped at 50 Hz, to remove the interference of industrial frequency. Data acquisition was performed using the international 10-20 modified 64-channel EEG cap with ground electrode at the midpoint of Fz and FPz. In offline experiments, EEG signals were acquired using 60 electrode channels (excluding M1, M2, CB1, CB2), with left posterior mastoid as the reference electrode. On the other hand, the online experiments used 9 electrode channels (Pz, Oz, O1, O2, POz, PO3, PO4, PO5, PO6) to record the EEG data, which was then sent from data acquisition side to presentation side via TCP/IP protocol, followed by the realtime feedback analysis. Visual stimuli program was written using the Psychtoolbox of MATLAB software.
B. Experimental Design 1) Paradigm Design: In our experiments, the stimulus interface presentation device was a 1920 × 1080 pixels LCD monitor with a refresh rate of 60 Hz. The stimulation interface is presented in Fig. 1(a), where the interface has 9 rectangular boxes (each with 180 × 80 pixels) distributed evenly in a 3 × 3 format, and the strips with variable luminance and position were set in the center of rectangular boxes as the stimulus targets. The size of strip was 20 × 80 pixels, and the distances between adjacent rectangular boxes in the horizontal and vertical directions were 460 pixels and 280 pixels, respectively.
In this work, the sinusoidal sampling stimulus encoding method was adopted to realize the stimulus frequencies and phases. Meanwhile, the luminance change was realized using the joint frequency-phase modulation (JFPM) method. Essentially, we can change the luminance of targets via setting the stimulus sequence s ( f, ϕ, i) corresponding to the frequency f and phase ϕ of stimulus with the following equation as: where i indicates the frame index in a stimulus sequence and R indicates the screen refresh rate. The dynamic range of stimulus sequence s ( f, ϕ, i) is from 0 to 1, where 0 represents dark and 1 indicates maximum luminance. In addition, Fig. 1(b) displays the frequency and phase values for each strip. The flicker frequencies of targets change from 8 Hz to 12 Hz with an interval of 0.5 Hz, while the interval of phase is 0.5π .
The initial position of each strip was set in the center of the rectangular box, and the instantaneous position of each strip l (F, i) can be obtained by the equation given as follows: where A is the displacement of sinusoidal trajectory, and F indicates the horizontal motion frequency. In this study, A was set to 80, and the dynamic range of strip's position was from −80 pixel to 80 pixel. Moreover, the horizontal motion frequencies were 0 Hz, 0.2 Hz, 0.4 Hz and 0.6 Hz. Fig. 1(c) illustrates the visual stimulation modulated simultaneously by luminance and motion.
2) Experiment Procedure: The aim of offline experiment was to optimize the BCI system parameters. In particular, the offline experiment consisted of 24 blocks, and included four different horizontal motion frequencies (i.e., 0, 0.2, 0.4, and 0.6 Hz). Thus, each frequency condition corresponded to 6 blocks in the offline experiment. Besides, each block included 9 trials, and each trial lasted for 5 s. Each trial started with a red strip that lasted for 0.5 s to prompt a target strip, and then all the strips began to flicker simultaneously on the screen for 4 s. After the flashing task, all targets returned to the center of rectangle box and stopped flashing for 0.5 s, which concluded the trial. Furthermore, all the subjects were asked to perform the NASA-TLX questionnaire [28] and the Comfort Level scale [29] after the offline experiments.
The goal of online experiment was to validate the feasibility of the proposed novel stimulus encoding method. The validation experiment included two conditions (i.e., horizontal motion frequency of 0 Hz and 0.2 Hz). Here, each condition consisted of 20 blocks, with each block having 9 trials. For the 0 Hz horizontal motion frequency condition (i.e., stationary visual stimuli), each trial included 3 s of flashing stimulation and 0.5 s of gaze-shifting, thereby lasting for a total of 3.5 s. Whereas, for the 0.2 Hz horizontal motion frequency condition, each trial included 4 s of visual stimulation and 0.5 s of gaze-shifting, with a total time span of 4.5 s. After the stimulus offset, visual feedback (i.e., if the identification result of SSVEP signal was consistent with the prompt, there was a beep, else there was no beep) was provided to the subjects in real-time. Similar to offline experiments, all subject were required to perform the NASA-TLX questionnaire and the Comfort Level scale after the online experiment.
C. Signal Processing 1) Amplitude of SSVEPs: Fast Fourier transform (FFT) in time domain was computed for the SSVEP data with length of 4s obtained in the offline experiments, to calculate the SSVEP amplitude spectrum of the Oz channel in all trails.
2) Decoding Algorithm of SSVEPs: FBCCA decoding method [16] can effectively utilize different spectral characteristics of the fundamental and harmonic components of SSVEPs, thus improving the recognition accuracy of SSVEPs. Accordingly, this study used FBCCA method to decode the SSVEP signals. The FBCCA method mainly includes filter bank analysis, canonical correlation analysis (CCA) processing, and target recognition. Firstly, different band-pass filters were used to divide the SSVEP into multiple subband signals (X S B n , n = 1, 2, . . . , N ). Secondly, the typical correlation coefficients between each sub-band component and each flicker frequency f k corresponding to sine and cosine reference signal (Y f k , k = 1, 2, . . . , 9) were calculated. Here, correlation vector ρ k consisted of N correlation coefficients corresponding to kth flicker frequency f k .
The reference signal Y f k is related to f k as: where N h indicates the number of harmonics, N p is the number of sampling points and f s is the sampling frequency. Then, the correlation vector is: where ρ (x, y) represents the correlation coefficient of x and y. The square of each typical correlation coefficient multiplied by the corresponding weighting factor forms the featureρ for target recognition.ρ where i indicates the number of sub-bands. Next, the weight coefficient w(n) corresponding to correlation coefficient of each sub-band can be expressed as follows: where i is the index of sub-bands. Based on a previous study [16], a and b used in this study were set to 1.25 and 0.5, respectively. And N h was set to 5 in this study. Notably, the frequency of reference SSVEP signal corresponding to the maximum correlation coefficient is the frequency of target stimulus.

D. System Performance Evaluation
Information transfer rate (ITR) is a widely used metric to evaluate the BCI performance, which can be calculated by following equation.
where N denotes the number of targets, P denotes the recognition accuracy rate, and T is the time required to output a single command. In this work, the calculation of ITR included the gaze-shifting time of 0.5 s.
E. Subjective Feelings Assessment 1) NASA-TLX Questionnaire: The NASA Task Load Index (NASA-TLX) scale is widely used to assess the subjective workload [30]. Likewise, this study also recorded the NASA-TLX scores of each subject. The six sources of workload indicated by the questionnaire are: mental demand (MD), physical demand (PD), time demand (TD), performance (PE), effort (EF), and frustration (FR). These six workload-related factors were compared in a two-by-two fashion, yielding a total of 15 pairs. Subjects were asked to select a factor in each pair that acted as the most dominant source of workload variation in these tasks, and the number of times each factor was selected was counted as its weight. Weights of factors are represented by w1, w2, w3, w4, w5, w6, ranging from 0 to 5. Then, each related factor was scored for accuracy on a scale of 0-100. The final NASA-TLX score was the average of the weighted sum scores of accuracies for these 6 elements. A lower NASA-TLX score means that the subject can perform the task more easily. (9), shown at the bottom of the next page.
2) Comfort Level Scale: In this study, the feedback on comfort level of each experimental condition was also provided by each subject. The subjective assessment questionnaire from a previous study [29] was adopted, where the subjects were required to grade each experimental condition with a 6-point scale ranging from 1 (totally unacceptable) to 6 (a good experience). Fig. 2 shows the fundamental SSVEP amplitude topographies and mean SSVEP amplitude spectra for each horizontal motion frequency condition. As shown in Fig. 2, the topographies of fundamental SSVEP signals induced by the four different horizontal moving speeds are similar and the strong SSVEPs are mainly obtained at the parieto-occipital electrodes. Furthermore, spatial distributions of fundamental SSVEPs induced by nine different stimulus frequencies also reveal that SSVEPs are stronger at parieto-occipital area. In particular, the amplitude of SSVEPs at channel Oz is higher than other channels at parieto-occipital area, thus Oz channel was chosen to calculate the SSVEP amplitude spectrum. Furthermore, all four conditions show significant peaks at both fundamental and harmonic response frequencies, which indicates that the visual stimulus of all four motion frequencies in this paradigm can elicit stable SSVEP. Moreover, the amplitude of each harmonic decreases with the increasing number of harmonics, and the peak value of SSVEP in the moving state at each harmonic is lower than that in the stationary state. Fig. 3 shows the mean SSVEP amplitude values of each flicker frequency for four horizontal motion frequency conditions and the mean SSVEP amplitude values of each motion frequencies, respectively. Fig. 3(a) indicates that stationary paradigm (i.e., 0Hz motion frequency condition) outperforms moving paradigm in eliciting the SSVEPs except for the conditions of three flicker frequencies (i.e., 8 Hz, 9.5 Hz, and 11.5 Hz), and there is an overall decreasing tendency of the SSVEP amplitude as the horizontal motion frequency increases although some exceptions exist (i.e., the condition at 8.5 Hz and 9.5 Hz). A two-way repeated measure ANOVA, with horizontal motion frequency (4 levels) and flicker frequency (9 levels) within subjects (17 subjects) as variables of the SSVEP amplitudes, only showed a significant main effect of horizontal motion frequency ( p<0.05). Pairwise comparisons demonstrated that the SSVEP mean amplitude under 0 Hz horizontal motion frequency condition was notably higher than the amplitudes under other conditions (all p<0.05). And the SSVEP mean amplitude under 0.4 Hz horizontal motion frequency condition was not significantly different from the SSVEP mean amplitude under 0.6 Hz horizontal motion frequency condition ( p>0.05). Subsequently, the levels of significant differences between four motion frequency conditions have been marked in Fig. 3(b). One-way repeated measure ANOVA was performed for each motion frequency conditions at different flicker frequencies, and there existed significantly difference among motion frequency conditions at most flicker frequency conditions ( p<0.05). Fig. 4 shows the BCI performance under the four horizontal motion frequency conditions with different data lengths. Evident from Fig. 4, BCI performance decreases as the horizontal motion frequency increases, where horizontal motion frequency of 0 Hz resulted in the highest BCI performance. Moreover, the highest ITR (i.e., 39.94 ± 15.25 bits/min) is obtained at a data length of 2.5 s, however at this length, the corresponding recognition accuracy is only 80.72 ± 15.69 %. As shown in Fig. 4(b), recognition accuracy increases with the data length, and the recognition accuracy and ITR under 0 Hz horizontal motion frequency condition are 84.97 ± 14.04 % and 37.98 ± 12.62 bits/min, respectively, at 3 s of data length. Additionally, paired t-test showed that ITR obtained at 2.5 s data length is comparable to the ITR obtained at 3 s data length ( p>0.05). Thus, a data length of 3 s was chosen to build the subsequent online BCI system based on 0 Hz horizontal motion frequency. Moreover, as illustrated in Fig. 4(b), the recognition accuracy under 0.2 Hz horizontal motion frequency condition is just over 80% (i.e., 80.5 ± 12.27 %) at 4 s data length. Therefore, a data length of 4 s was chosen to build the subsequent online BCI system based on 0.2 Hz horizontal motion frequency. As for 0.4 Hz and 0.6 Hz moving conditions, the recognition accuracies of 4 s data length are below 80 %, thus we did not build online experiment for above two moving conditions. Fig. 5 redraws the average recognition accuracy of 3 s and 4 s data length in Fig. 4(b) under four horizontal moving speed conditions and the univariate regression analysis was carried out. It intuitively shows that average recognition accuracy decreases linearly as the horizontal motion frequency increases.

A. Offline Experiment Results
Furthermore, the average NASA-TLX scores at the four different horizontal motion frequency conditions were 75.24 ± 7.86, 74.12 ± 10.75, 75.65 ± 11.20, 76.24 ± 10.46, respectively. Although a one-way repeated measure ANOVA on NASA-TLX score revealed that there was no significant difference among the four experimental conditions ( p>0.05), the 0.2 Hz horizontal motion frequency condition obtained the minimum NASA-TLX score. Besides, the average scores of Comfort Level scale were 5.12 ± 0.78, 5.35 ± 0.70, 5.06 ± 0.90, 4.35 ± 1.11, respectively, for four frequency conditions. A one-way repeated measure ANOVA on comfort scores showed that there exists a significant difference among the four horizontal motion frequency conditions ( p<0.05). Post hoc analyses demonstrated that the comfort score under 0.6 Hz horizontal motion frequency condition was notably lower than the scores under other conditions. Interestingly, this result in fact indicates that a high horizontal motion frequency may reduce the comfort of visual stimuli. The 0.2 Hz horizontal motion frequency condition exhibited the maximum comfort score. Therefore, the visual experience under the 0.2 Hz horizontal motion frequency condition was optimal due to the lowest task load and the highest subjective comfort level for subjects in this condition.

B. Online Experiment Results
According to above mentioned offline analysis, a 3-s stimulation duration and a 4-s stimulation duration were  used in the 0 Hz horizontal motion frequency condition and the 0.2 Hz horizontal motion frequency condition, respectively. Both the two conditions adopted a 0.5-s gazeshifting time. Thus, the time for the 0 Hz horizontal motion frequency condition and the 0.2 Hz horizontal motion frequency condition to output one command was 3.5 s and 4.5 s, respectively. Table I shows the BCI performance recorded in the online experiments. For the 0 Hz horizontal    The average NASA-TLX scores for the 0 Hz and 0.2 Hz horizontal motion frequency conditions were 85.08 ± 3.85 and 85.08 ± 6.65, respectively. Here, paired t-test revealed no obvious difference between the NASA-TLX scores of two experimental conditions ( p>0.05). On the other hand, the average comfort scores for the 0 Hz and 0.2 Hz horizontal motion frequency conditions were 4.42 ± 0.79 and 4.67 ± 1.15, respectively, and the corresponding paired t-test also revealed no obvious difference between the comfort scores of two experimental conditions ( p>0.05). The results of the questionnaire and scale were indeed consistent. It is worth noting that although the trends of the two scale scores are generally consistent with those in the offline experiment, the overall NASA score is higher while the overall comfort scale score is lower, presumably because there is only 0.5 s of gaze-shifting between the two trails in the online experiment, which in turn leads to a larger task workload and a corresponding decrease in system comfort.
Since 9 participants joined both offline and online experiments, we compared the BCI performance of these 9 participants in the offline and online experiments. For the 0 Hz horizontal motion frequency condition, the average recognition accuracies of offline and online experiments were 86.42 ± 15.07 % and 86.73 ± 5.86 %, respectively. For the 0.2 Hz horizontal motion frequency condition, the average recognition accuracies of offline and online experiments were 82.10 ± 10.19 % and 82.84 ± 11.28 %, respectively. Besides, paired t-tests confirmed that the BCI performance of offline experiment was similar to that of the online experiment under two moving conditions (all p>0.05). In addition, subjective scale results in offline and online experiments were also comparable, showing no significant difference ( p>0.05). These well-matched results thus prove the stability of the proposed experimental paradigm.

IV. DISCUSSION
This study proposed a novel stimulus encoding method based on simultaneous modulation of luminance and motion. A 9-target SSVEP-BCI was built using the proposed stimulus encoding method, and offline and online experiments verified the feasibility of the proposed BCI system.
Most of the existing SSVEP-BCI studies adopt the stationary visual stimulus to elicit the SSVEPs, while only a few studies have used moving visual stimulus. The present study verified that the BCI performance decreases with the increase of horizontal motion frequency, which is coherent with the results of previous studies [19], [20]. For example, Kanoga et al. [19] investigated the effect of head movement on the SSVEPs and found that the frequency recognition accuracy decreases with the increasing head movement speed. Duan et al. [20] evaluated the effect of moving visual stimulus on the SSVEP-BCI performance and observed that the moving paradigm did not lower the SSVEP-BCI performance, thereby verifying the feasibility of moving paradigm. Similarly, the present study validates the feasibility of horizontal moving paradigm. Our online experimental results show an average recognition accuracy of 83.15 ± 9.88 % under the 0.2 Hz horizontal motion frequency condition, which suggests that although the moving visual stimulus can affect the SSVEP-BCI performance in general, it can also be used to design and implement SSVEP-BCIs. Furthermore, unlike previous studies where luminance and motion change were used to encode different regions of the target [31], we designed the whole target to move horizontally while the luminance changes. In this case, subjects were asked to follow the moving target during the task time instead of staring at a fixed center point of the target. No obvious intermodulation frequency component was observed in the present study. An additional difference in this work compared to above mentioned previous studies is the further reduction in the area of the visual stimulus. The above two studies adopted a square visual stimulus, where in the study of Kanoga et al. [19], the stimulus target was a 5 × 5 cm square. Meanwhile, the visual stimulus was a 120 × 120-pixel square in the study of Duan et al. [20]. In contrast, a smaller visual stimulus was used in this study (i.e., a 20 × 80-pixel strip). On one hand, a small visual stimulus facilitates the presentation of a larger number of targets on the monitor. On the other hand, small visual stimulus may be more comfortable to subjects [29]. It should be noted that largesized flicker stimulus tends to induce the visual fatigue in SSVEP-based BCIs [29], and the mean comfort score of the proposed paradigm exceeded 4, indicating that the proposed paradigm is visually acceptable [29]. In addition, the system with 0.2 Hz horizontal motion frequency provided the best visual experience for subjects according to the results of NASA-TLX questionnaire and the Comfort Level score. These results indicated that the system with 0.2 Hz horizontal motion frequency is expected to develop a more comfortable BCI system.
Further research based on the simultaneous modulation of luminance and motion investigated in this study can be developed from the following perspectives. Firstly, in terms of the algorithm for decoding the SSVEP, we elected the advanced untrained FBCCA method, on which certain improvements can be made to further enhance the system performance. For instance, Yang et al. [32], [33] designed a new dynamic window to improve the performance of SSVEP-BCI, which can automatically find the optimal data length to reach a higher ITR. In addition, trained methods such as TRCA and TDCA can also be considered. Progressively, Lin et al. [34] proposed transfer-extended Canonical Correlation Analysis (t-eCCA) method, and performance obtained by this method is comparable with that of TRCA and TDCA with less training time. At the same time, a transfer learning framework (ALPHA) was proved to outperform CCA and TRCA [35]. Secondly, the influence of flickering area of each stimulus target, distance between the targets, and other factors related to interface settings on the induced SSVEP [36] also needs to be further explored. In this study, in order to ensure that the subjects stare at the flickering targets without being disturbed by other targets (especially in the moving state), the stimulus targets are scattered. To balance between the target interval (affecting the degree of interference) and the gaze shifting time (affecting the acceptable response speed of subjects), we need to set up further comparison experiments with different parameters to build a BCI system under the optimal configuration. Thirdly, the present study adopted healthy subjects to test the proposed system. For future proposals, it is desirable to test the proposed system by the patients with impaired visual motion perception. Finally, in future work, we will monitor the head and eye motion sate of participants, and then further investigate whether the Electrooculogram signal or the EMG signal will affect SSVEP identification.

V. CONCLUSION
This study introduced a novel stimulus encoding method based on simultaneous modulation of luminance and motion, and then used it to build a 9-target SSVEP-based BCI. We compared the BCI performance at four horizontal motion frequencies. Offline experimental results showed that the BCI performance decreased with the increasing horizontal motion frequency. But low horizontal motion frequency still can be used for building BCI system. The online experimental results showed that the subjects achieved 85.00 ± 6.77 % and 83.15 ± 9.88 % accuracy for the superimposed horizontal periodic motion frequencies of 0 and 0.2 Hz, respectively. These results verified the feasibility of the proposed system.