Keywords

1 Introduction

The purpose of studying the characteristics of electroencephalographic (EEG) signals in the rifle shooting process is to develop an assistive tool that can assist the upper-limb disabled soldier to continue the firing task and provide a new direction for the manufacture of weapons. The key point of the tool is to analyze and extract the feature of the EEG signal that can regulate control system. Aiming is a very important part [1] in the process of shooting; the aiming performance determines the accuracy of shooting. Accordingly, the content of the data was of interest for the purpose of gaining a better understanding of the psychophysiological characteristics of an expert shooter during the aiming period and adopting signal processing approach to verify the feasibility of the assistive tool solution.

EEG recordings have been employed in studies of the aiming period in the shooting. Bouchard and Tetreault studied the visual impairment effects on the body stability at the time of aiming [2,3,4]. Joy Myint’s researched using auditory to aid defective shooter (vision impairment) to aim [5]. Combining the technologies of EEG and eye tracking has been advocated by Janelle et al. as a means of gaining a better understanding of the underlying mechanisms that regulate human visual attention [6]. Hatfield exposed the cortex activity of shooting experts and novices while aiming [7,8,9]. To better analyze cortical functioning in the shooting or archery and the “neural efficiency hypothesis”, Event Related Desynchronization/Synchronization (ERD/ERS) analysis has been used by researchers [10,11,12,13]. Vijn et al. [14, 15] proposed the concept of occipital EEG alpha power during the aiming period, and M. Doppelmayr studied the time course of frontal midline theta (Fmθ) during the aiming period in rifle shooting [16]. But few people extracted the feature of EEG signal when aiming.

In this paper, we analyzed the characteristics of the EEG from the rifle shooting experts during the aiming period to get the EEG feature which is the mark of the cerebral cortical activity and the brain region where the feature located. We used the wavelet transform to denoise and feature enhancement. Then the power of alpha, beta and theta were calculated. Support vector machines were used to classify the two classes of signals (the holding signal and the aiming signal), after extracting the EEG features.

2 Method

2.1 Participants

Ten excellent soldiers from the army volunteered to take part in the study. The shooters, who were aged 25–40 years, had a minimum of 4 years of shooting experience. All shooters were healthy men, had visual acuity or corrected visual acuity of 4.8 or more, held the rifle in their right hand and used their right eye for aim. Each participant had signed consent before taking part in the experiment.

2.2 Procedure

The participants were initially informed about the procedures and provided verbal and written instructions regarding the shooting task. After a warm-up period in a shooting range, the shooter was asked to perform three 50-shot runs to the target, using a standard 56-type semi-automatic rifle. The distance between the shooter and the target was 70 m. Each shot involved three periods: preparatory period, aiming period, the shot release period. Voice cures played at the beginning of each period with which the subject was instructed to perform the task. Figure 1 shows a 2-min for break time between the two runs, and the two trials interleaved with 1 s.

Fig. 1.
figure 1

Shooting experiment

Full size image

2.3 Data Collection

EEG data was recorded from 32 electrodes arranged according to the 10–20 system [17] in a BP cap (ElectroCap Inc, Florida, USA), and was made using BP ActiCHamp amplifiers. EEG data continuously recorded with 1000 Hz sampling frequency. The ground electrode and common reference were positioned at Cz to ensure low impedance values (generally below 5 KΏ). Eye-movement (EOG) artifact monitored with two electrodes attached superior to and on the external canthus of the right eye. An eye patch was worn over the left eye to minimizing muscle artifact in the EEG record due to squinting while aiming.

2.4 Data Analysis

EEG data were band-pass filtered between 0.1 to 45 Hz. After wavelet decomposing and reconstructing, the data were quantified in the Theta (4–8 Hz), low Alpha (8–12 Hz), and Beta (16–24 Hz) bands. We have used four parameters (the alpha bands power, the beta bands power, the theta bands power, the beta bands power and theta power (β/θ)) to compare the differences between the data of the holding period and the aiming period. Then the features were extracted and the two classes of the signals were classified by SVM.

2.4.1 Methods

Wavelet Transform.

The essence of the wavelet transform (WT) [18] is to decompose the signal into a series of wavelet functions, and they perform multiscale refinement analysis by stretching and translation. WT has the characteristics of multi-resolution analysis. It based on this characteristic that wavelet transforms better for those non-stationary signals, especially those with rapidly changing signals. In practice, the wavelet transform is effective in signal de-noising, weak signal extraction and signal singularity analysis.

Let \( x(t) \in L^{2} (R) \) (\( L^{2} (R) \) is the square integrable function space on R), then the formula of wavelet transform of x (t) as follows [19]:

$$ WT_{x} (a,\tau ) = \frac{1}{\sqrt a }\int {x(t)\phi^{*} (\frac{t - \tau }{a})dt = \left. {\left\langle {x(t),\phi_{a\tau } (t)} \right.} \right\rangle } $$
(1)

Where \( \phi (t) \) is a function called elementary wavelet or mother wavelet, a > 0 is the scale factor, τ reflects the displacement, the value can be positive or negative.

The equivalent frequency domain expressed as

$$ WT_{x} (a,\tau ) = \frac{\sqrt a }{2\pi }\int {x(w)\psi^{*} (aw)e^{jw\pi } dw} $$
(2)

The discrete wavelet transform written as

$$ WT_{f} (\frac{1}{{2^{j} }},\frac{k}{{2^{j} }}) = \left\langle {\left. {f,\psi_{j,k} } \right\rangle } \right. $$
(3)

Then,

$$ \psi_{j,k} (t) = \psi_{{\frac{1}{{2^{j} }},\frac{k}{{2^{j} }}}} (t) = 2^{{{\raise0.7ex\hbox{$j$} \!\mathord{\left/ {\vphantom {j 2}}\right.\kern-0pt} \!\lower0.7ex\hbox{$2$}}}} \psi (2^{j} t - k) $$
(4)

\( \left\{ {\left. {\psi_{j,k} } \right\}} \right._{j,k \in Z} \) is the Riesz basis of L2 (R).

Let the sampling rate of the discrete signal is fs, the original signal and the sub-band signal distribution relationship as follows:

$$ \left[ {0,\frac{{f_{s} }}{2}} \right] = \left[ {0,\frac{{f_{s} }}{{2^{L + 1} }}} \right] \cup \left[ {\frac{{f_{s} }}{{2^{L + 1} }},\frac{{f_{s} }}{{2^{L} }}} \right] \cup \ldots \cup \left[ {\frac{{f_{s} }}{{2^{2} }},\frac{{f_{s} }}{2}} \right] $$
(5)

The sub-frequency band signals corresponding to these sub-bands are: \( x_{L}^{a} (n),x_{L}^{d} (n), \ldots ,x_{2}^{d} (n),x_{1}^{d} (n) \), then

$$ x(n) = x_{L}^{a} (n) + \sum\limits_{j = 1}^{L} {x_{j}^{d} (n)} $$
(6)

Where L is the number of decomposition steps and \( x_{L}^{a} (n) \) is the low-pass approximation component, \( x_{j}^{d} (n) \) is the detail component under different scales.

Support Vector Machine.

Support Vector Machines (SVM) has a unique advantage in dealing with small sample learning problems [20]. It has many advantages, such as strict theory, strong adaptability, global optimization, high training efficiency and good generalization performance. The SVM derived from the optimal classification surface. The optimal classification surface is that the classification surface can not only divide the two types of samples correctly but also make the classification gap maximum. Let train sample set {(xi, yi), i = 1,2, …, n}, then, xi ∈ RN belong to class 1, denote yi as 1, otherwise denote yi as −1. The goal of learning is to construct a discriminant function that classifies the two types of test data as accurately as possible.

In the case of linear separability, there exists a hyperplane such that the training samples are completely separated, and the classification hyperplane is

$$ w * x + b = 0 $$
(7)

Where w is an n-dimensional vector and b is an offset. (xi, yi) in the training set D is satisfied:

$$ \left. {\begin{array}{*{20}l} {w * x + b \ge 1,y_{i} = 1} \hfill \\ {w * x + b \le - 1,y_{i} = - 1,i = 1,2, \ldots {\text{n}}} \hfill \\ \end{array} } \right\} $$
(8)

Where \( w * x \) is inner product. The above formula can abbreviated as:

$$ y_{i} (w * x_{i} + b) \ge 1,i = 1,2, \ldots ,{\text{n}} $$
(9)

If all the vectors in the training sample set are correctly partitioned by a hyperplane and the distance between the sample data closest to the hyperplane and the hyperplane is the largest, then the hyperplane is the optimal hyperplane, like Fig. 2.

Fig. 2.
figure 2

Optimal classification hyperplane

Full size image

The optimization of the hyperplane can transformed into the quadratic programming problem, and the training data satisfy the following formula:

$$ \left. {\begin{array}{*{20}l} {\mathop {\hbox{min} }\limits_{w,b} \frac{1}{2}||w||^{2} } \hfill \\ {y_{i} (w * x_{i} + b) \ge 1,i = 1,2, \ldots ,{\text{n}}} \hfill \\ \end{array} } \right\} $$
(10)

In this paper, two firing states are separated by SVM. The appropriate kernel function of SVM can make the accuracy of classification optimal. We use the Soft Spaced Support Vector Machine (SVM) whose kernel function is radial basis function (RBF).

2.4.2 Data Preprocessing

To weaken noise and enhance feature in the signals of aiming electroencephalogram data, the independent component analysis (ICA) were applied to identify and remove any remaining artifacts (e.g., eye blinks). The data acquired had a total amplitude of less than 100 uV. They were band-pass filtered between 0.01 to 45 Hz, segmented into single epochs of 2 s duration, with each epoch starting at −2 s with respect to t = 0 (i.e. the instant when the shot was released). The holding data were also 2 s of each epoch. During the aiming period, the blink of right eye leads to a larger signal artifact in the forehead region, so the signal of FP2 channel had a strong noise. Finally, we dropped the channels in the brain’s frontal region (F3, Fz, F4), frontal-center region (FC5, FC6), central region (C3, Cz, C4), parietal region (P3, Pz, P4) and occipital region (O1, Oz, O2). After Preprocessing, we got two data sets XA and XB who stored in the form of ‘trials × time × channels’. Their size both were 1500 × 2000 × 14.

2.4.3 Data Analysis Process

The collection process of the signal shows the number of observed signals is less than the independent signals. For this reason, the sub-signals of alpha, beta and theta obtained by multi-resolution analysis of the wavelet transform. The data of holding period and aiming period were decomposed and reconstructed using db6 wavelet function as wavelet basis. Since the ‘Daubechies’ wavelet has smoothing property and the total entropy value obtained by using db6 was the smallest. The power of the sub-signals of alpha bands, beta bands, theta bands and beta/theta calculated separately, and the data format saved as 14 × 1500 × 4. Figure 3 shows the signals of alpha, beta and theta of the F3 channel in the holding period and the aiming period.

Fig. 3.
figure 3

The holding period and aiming period, the signal of alpha, beta and theta at the F3 channel

Full size image

Alpha, beta, theta spectral power and beta/theta in the five regions of the brain were subjected to separate 2 × 14 × 1500 × 4 (shooting period × channels × time × parameters) Paired T-test with repeated measures on channels. The parameters with significant changes used as input vector of SVM and the classification results of two classes at channel O1 and O2 obtained. In this paper, the kernel function of SVM was Gaussian radial basis function, and the marginal factor were 25, 35, 45 respectively.

3 Results

During the aiming period and the holding period, the results obtained by pairwise T-test of the four parameters (α, β, θ, β/θ) in the 14 channels (F3, Fz, F4, FC5, FC6, C3, Cz, C4, C3, Cz, C4, O1, Oz) are shown in Table 1. We inferred statistical significance levels (p < 0.05 and p < 0.01) from the 95th and 99th percentiles of the distribution.

Table 1. The results of the paired T-test
Full size table

From the summary of group mean EEG power in Table 1, we can see that statistical analysis revealed several significant differences in power for the frontal region. In detail, during the holding period and the aiming period, the power of α existed obvious differences (p < 0.01) in frontal, frontal- central, central (C3, Cz), parietal, occipital region. Similarly, there was a main effect (p < 0.01) for the β power in the region of frontal, central, parietal, occipital, with significant changes (p < 0.05) in the frontal and parietal region for θ power. Meanwhile, the θ/β was significant differences (p < 0.05) in frontal, frontal- central, occipital region.

The results of power analysis in the alpha band and beta band shown in Fig. 4, during the aiming period, the power increased significantly in both the left and right hemispheres of the brain, especially in the occipital and apical regions. Figure 5 shows power in the left hemisphere of the brain to increase more than the right. It can be seen from Fig. 6 that there was a significantly reduced at the F3 site for the theta power, with obviously increased at the F4 and Cz site.

Fig. 4.
figure 4

Alpha and beta power in the holding period and aiming period

Full size image
Fig. 5.
figure 5

The alpha and beta power in the left and right hemispheres of the brain during the aiming period

Full size image
Fig. 6.
figure 6

Theta rhythm in holding and aiming state of the power value

Full size image

The combined of alpha power and beta power as the input of SVM and SVM was used to calculate the classification accuracy of the two classes signals (the holding signals and the aiming signals) at O1, O2, Oz site. Similarly, theta was used as input to the SVM to compute the classification accuracy (shown in Table 2) of the two classes signals at Fz site. In the data set, 1200 trials data as training data, and other 300 trials as the test data.

Table 2. The accuracy at Fz, O1, O2 and Oz channel with different marginal factor
Full size table

4 Discussion

The comparison of the cortical activity patterns obtained from the marksmen during the holding period and the aiming period of the target shooting task revealed that the differences observed in each of the frequency bands assessed. Based on the activation patterns it seems plausible to conclude that the eye, especially the right eye, plays a dominant role during the aiming period. This deduction based on the finding of increased alpha power and beta power in the parietal and occipital regions of the experts during the aiming period of the target shooting task and the power on the left hemisphere of the brain was greater than the right hemisphere which can be seen in Fig. 5. This result was similar to that reported previously [21,22,23]. Additional insights regarding the psychological processes of the expert performers were enabled by examination of spectral content beyond that typically reported in the sports psychophysiology literature. Such as, Doppelmayr et al. discussed the frontal midline theta (Fmθ) could be observed in the human electroencephalogram (EEG) at the frontal midline electrode location Fz before shot release [16, 24]. Smith et al. [25] described this spectral component recorded at the midline frontal area as indicative of sustained attention. Such a finding also showed in Fig. 6 and in which can be seen increased theta power in the frontal region (F3, F4). The increased theta power observed at channel Cz means that the central region is activated, but little has been previously reported and requires further validation. Lutsyuk et al. [26] reported θ/β represented the level of attention. From Table 1, we can see that the value of θ/β in the frontal, apical and occipital regions were significantly changed, indicating that the marksmen possess a higher degree of attentional focus.

During the aiming period, the main cognitive task of the shooter was to stare at the target and the attention intensification. One could deduce that the alpha power and beta power in the occipital region (O1, O2, Oz) were the feature vectors of the EEG during the gazing period and theta power in the frontal region (Fz) was the EEG feature vector during focused concentration. It can be seen from Table 2 that the SVM can separate the two classes of signals (the holding signal and the aiming signal) in above the four cases. The optimal classification results were obtained by combining alpha power and beta power as EEG feature at the channel of O1. In the three cases (marginal factor was 25 or 35 or 45), when the marginal factor was 25, the optimal classification accuracy was obtained.

5 Conclusion

The EEG signal of excellent shooters during the aiming period of rifle shooting were analyzed by power spectrum method. The power of alpha bands and beta bands increased significantly in the occipital and apical region, and theta bands increased the power at F3 and F4 channel. Both alpha-power and beta power combination, or theta power as EEG feature, support vector machine can separate the two classes of signals (the signal of holding and aiming). As can be seen in Table 2, the optimal classification accuracy was obtained at the O1 channel when the marginal factor was 25. The further study will continue to analyze the characteristics of EEG signals during aiming and firing, and look for better classification algorithms to improve the accuracy of classification and prepare for the control system.