Online Detection of P300 related Target Recognition Processes During a

Demanding Teleoperation Task

Classiﬁer Transfer for the Detection of Missed Targets

Hendrik Woehrle

and Elsa Andrea Kirchner

1,2

Robotics Innovation Center, German Research Center for Artiﬁcial Intelligence (DFKI GmbH)

Robert-Hooke-Str 5, Bremen, Germany

Robotics Lab, University of Bremen, Robert Hooke Str. 5, Bremen, Germany

Keywords:

Brain Computer Interfaces, Embedded Brain Reading, P300, Single Trial Detection, Exoskeleton.

Abstract:

The detection of event related potentials and their usage for innovative tasks became a mature research topic

in the last couple of years for brain computer interfaces. However, the typical experimental setups are usually

highly controlled and designed to actively evoke speciﬁc brain activity like the P300 event related potential. In

this paper, we show that the detection and passive usage of the P300 related brain activity is possible in highly

uncontrolled and noisy application scenarios where the subjects are performing demanding senso-motor task,

i.e., telemanipulation of a real robotic arm. In the application scenario, the subject wears an exoskeleton to

control a robotic arm, which is presented to him in a virtual scenario. While performing the telemanipulation

task he has to respond to important messages. By online analysis of the subject’s electroencephalogram we

detect P300 related target recognition processes to infer on upcoming response behavior or missing of response

behavior in case a target was not recognized. We show that a classiﬁer that is trained to distinguish between

brain activity evoked by recognized task relevant stimuli and ignored frequent task irrelevant stimuli can be

applied to classify between brain activity evoked by recognized task relevant stimuli and brain activity that is

evoked in case that task relevant stimuli are not recognized.

1 INTRODUCTION

Online-analysis and detection of speciﬁc patterns in

electroencephalographic (EEG) data has been used

for various applications, e.g., brain computer inter-

faces (BCIs). Current EEG-based BCIs are using

classiﬁcation and data dependent signal processing

methods to detect the patterns in the EEG. Therefore,

they highly depend on training data that has to be

used for the calibration of the system before they can

be used to detect the patterns in the application data.

Usually, the training data has to be subject-speciﬁc,

i.e., it has to be acquired from the subject in train-

ing sessions directly before the usage of the system.

Further, the recorded data should be clean, i.e., free of

artifacts that might affect the training or detection pro-

cess, as well as task speciﬁc, i.e., must consist of data

that is directly related to the patterns that are supposed

to be detected. Therefore, most results are conducted

in highly controlled artiﬁcial scenarios, where most of

the possible disturbance sources have been excluded

by experimental design and the subject may even be

ﬁxed in a speciﬁc position.

For many applications this is no draw back, espe-

cially if BCIs are applied as active interfaces, i.e., to

control a machine or computer (Farwell and Donchin,

1988; Guger et al., 1999; Wolpaw et al., 2002; Reud-

erink, 2008; Nijholt et al., 2008). If, however, the pat-

terns that should be detected in the brain activity are

no longer actively produced, as it is the case for pas-

sive BCIs (Zander et al., 2010; George and L

ecuyer,

2010), then background EEG that is evoked by the

active task may overlay with the relevant brain. Since

the subject is possibly performing different tasks, the

background EEG may differ strongly depending on

the situation and performed active task or action of

the user and thus affects the training data.

For future practical applications of passive ap-

proaches (Zander et al., 2010; George and L

ecuyer,

2010; Kirchner et al., 2010; Haufe et al., 2011; Kirch-

ner and Drechsler, 2013), it is required to expose

the systems and subjects to concrete, realistic use-

cases, that are more uncontrolled and performed in

perturbed environments. These conditions likely in-

Woehrle H. and Kirchner E..

Online Detection of P300 related Target Recognition Processes During a Demanding Teleoperation Task - Classiﬁer Transfer for the Detection of Missed

Targets.

DOI: 10.5220/0004724600130019

In Proceedings of the International Conference on Physiological Computing Systems (PhyCS-2014), pages 13-19

ISBN: 978-989-758-006-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

crease the amount of noise in the training as well as

application data and may therefore impair the detec-

tion accuracy.

A further problem exists, if the amount of training

data is small. It might not be possible to acquire a

large amount of training examples in complex appli-

cation scenarios. This is a general problem, since the

detection accuracy of data-dependent signal process-

ing and classiﬁcation methods depends on the amount

of available training data. Hence, approaches that can

handle a reduced amount of training examples must

be developed and applied.

One approach is to transfer a classiﬁer between

classes, i.e., to perform a classiﬁer transfer (Pan and

Yang, 2010). It could be shown that classiﬁer transfer

for the detection of patterns in the EEG performs well

for a transfer of classiﬁer between tasks in which the

same event related activity has to be detected (Itur-

rate et al., 2013) or between similar types of event

related potentials (ERPs) like different types of error

potentials (Kim and Kirchner, 2013). In a recent work

we showed that the transfer of classiﬁer is also pos-

sible between classes that ”miss” a pronounced pat-

tern, i.e., the P300 (Kirchner et al., 2013). Hence,

the data processing methods (classiﬁers and spatial

ﬁlters) need not to be trained and tested on examples

that are evoked by the same brain processes, like same

or similar error detection processes, but by brain pro-

cesses that evoke brain pattern, which are similar in

shape and characteristics, i.e., miss a prominent ERP

or pattern of ERPs.

By now, our investigations have been conducted

in controlled experimental setups in an ofﬂine fash-

ion. In this paper, we investigate the ability to detect

the P300 ERP in a demanding dual task application

scenario that combines an oddball paradigm with a

second task. We show that the detection of P300 re-

lated target recognition processes and even more im-

portant the missing of target recognition processes can

be performed online while a subject is performing a

demanding and realistic interaction task that occupies

the operators attention. This task consists of the tele-

operation of a real robotic arm through a labyrinth via

a virtual immersion scenario.

The paper makes the following contributions: 1)

we demonstrate that the online, single trial detection

of the P300 potential is possible in an application

scenario that is affected by a high number of noise

sources and artifacts and requires dual task perfor-

mance from the subject (Kirchner and Kim, 2012),

i.e., distracts the subject from the perception of task

relevant stimuli; 2) we show that the few number of

examples of training data of a speciﬁc class can be

compensated to a certain degree by classiﬁer transfer.

2 APPLICATION SCENARIO

In the proposed application scenario, we investigate

whether it is possible to reliably detect target recogni-

tion processes as well as the missing of target recog-

nition processes while a subject is performing a de-

manding teleoperation task.

Precisely, the experimental setup was as follows

(see Fig. 1): The subjects were wearing an exoskele-

ton that covered their back and right arm (Folgheraiter

et al., 2012), and a smart glove on their hand that were

used as input devices for the teleoperation task.

In addition, participants were equipped with a

head mounted display (HMD) on which the teleopera-

tion site (including surroundings, labyrinth and robot)

could be seen in 3D. Additionally to the 3D environ-

ment, information from the control system, a camera

picture of the real scene and tools like a gyroscope de-

picting the orientation of the end-effector were at any

time in the operators ﬁeld of view. Head and hand

movements of the operator were tracked (InterSense,

Billerica, USA) and used to update the HMD.

The subjects had two main tasks that had to be

performed at the same time: a) to control a robotic

arm (teleoperation task) using the exoskeleton, and

b) to respond to speciﬁc messages (oddball task).

2.1 The Teleoperation Task

In the teleoperation task, the end-effector of a robotic

arm had to be steered through a labyrinth (see

Fig. 1 C). This task is similar to a wire loop game,

i.e., a certain path has to be followed and touching

the labyrinth had to be avoided. The movements of

the robotic arm were controlled via the exoskeleton

by mapping the state and relative position of the ex-

oskeleton components to a Mitubishi PA-10 robotic

arm (see Fig. 1 A in the lower right corner) via a vir-

tual model (see Fig. 1 A in the upper right corner)

thereof (depending on the concrete type of investiga-

tion, see Sec. 3).

The teleoperation task is difﬁcult and demanding

for the subject, and therefore forces the subject to con-

centrate on it. Further, the subject was requested to

rest from time to time. In each run 24± 8 rest periods

had to be performed (Seeland et al., 2013). During

rest the active exoskeleton kept the operators arm in

position. While this was the case the operator was not

allowed to respond to any warning (infrequent task

relevant stimuli, see below) that were presented to

him in an oddball fashion throughout the run.

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

Figure 1: Teleoperation scenario: (A) An operator controls a robot via an exoskeleton and a virtual environment; (B) The

operator monitoring system (OMS) integrates all important informations, it supervises brain states detected by brain reading,

monitors the events in the scenario and responses of the operator on target events; the allowed response time is adapted with

respect to brain state, i.e., whether a warning was recognized or not; (C) The operator is controlling the robot in the virtual

environment through an labyrinth; (D) The operator is interacting with the virtual environment to respond to warnings by

touching one of the response cubes.

2.2 The Oddball Task

In the oddball-task, the subject had to respond to in-

frequent task relevant stimuli (”targets”, see Fig. 2 B)

by touching one of three virtual cubes that were inte-

grated into the virtual scenario as response targets for

answering speciﬁc messages in a certain time frame

(see Fig. 1 D). Besides the task relevant messages

also frequent task irrelevant messages (”standards”,

see Fig. 2 A) were presented to the operator but re-

quired no response. Due to the oddball design (Polich,

2007), i.e., the presentation of infrequent task relevant

stimuli mixed with frequent task irrelevant stimuli, it

was expected that P300 related brain processes (Ku-

tas et al., 1977; Salisbury et al., 2001; Kirchner et al.,

2009; Polich, 2007; Kirchner and Kim, 2012) will

be evoked in case of recognized infrequent task rel-

evant messages but not in case of frequent task irrele-

vant messages or task relevant but not recognized, i.e.,

missed, task relevant messages (”missed targets”).

In recorded training data it was determined whether

a target was recognized or not by the occurrence or

missing of a response 10 sec after a target stimulus

was presented.

2.3 The Operator Monitoring System

To support the operator in the scenario, an operator

monitoring system (OMS) (Kirchner and Drechsler,

2013) was included into the setup. The purpose of

the OMS was to monitor the operators cognitive state

and the current state of scenario in order to adjust the

course of events that were shown to the operator to

minimize distraction of the operator and optimize her

or his support by appropriate scheduling of messages.

The allowed response time was 2 sec in case that

OnlineDetectionofP300relatedTargetRecognitionProcessesDuringaDemandingTeleoperationTask-Classifier

TransferfortheDetectionofMissedTargets

Figure 2: The warnings and responses shown in the virtual immersion teleoperation operator monitoring scenario: (A)

The frequently shown standard marker has the text STATE OK. (B) One of the three possible targets. The possibilities are

MAN ENTERED, OIL TEMPERATURE or COM REQUEST. (C) The operators response must be related to the warning. In

this case, it is the cube with the label COM. In case of a correct response, it is highlighted in green. (D) If the operator did not

respond in time, a repeated and highlighted second target is shown. (E) If the operator did not respond to the second target in

time, a third obtrusive error message is shown. (F) In any case it might happen that the operator touches a cube with a label

that does not correspond to the shown target. In this case, the cube is highlighted in red.

target recognition processes could not be detected af-

ter a target was presented (in case of missed target, see

Fig. 1 B, ﬁrst warning for ”oil check” was missed) or

to extend the allowed response time to 10 sec in case

target recognition process were detected after a target

was presented and recognized (in case of target, see

Fig. 1 B, second warning for ”oil check” was recog-

nized).

By adapting the allowed response time in this

manner allows to give the operator a longer time for

responses in case he recognized the warning, which

was especially relevant in case that the operator was

in a rest position and was not able to respond. On the

other hand a task relevant message could be repeated

rather quick in case that target recognition processes

could not be detected after a target message was pre-

sented.

To enable this adjustment of the virtual scenario

by the OMS, we used machine learning techniques to

detect the P300 ERP in the subjects EEG. Therefore, a

classiﬁer had to be trained to distinguish examples of

the class target from examples of the class missed tar-

gets online. Since the operators were highly trained in

the scenario they usually miss only a few target mes-

sages, hence the amount of missed target examples

that could be recorded was very low. Thus, we used

EEG activity evoked by irrelevant standard messages

instead of EEG activity evoked by missed target mes-

sages during the training phase of the data processing

to later distinguish between targets and missed tar-

gets.

3 METHODS AND

EXPERIMENTAL

PROCEDURES

The experiments were performed with three male sub-

jects (age 27.33 ± 2.52), with a total of ten recording

sessions.

3.1 Setup and Data Acquisition

The data was acquired with a 64-channel actiCap sys-

tem and two ampliﬁers (both from Brainproducts,

Munich, Germany) at 5 kHz sampling rate. Four

electrodes (FC5, FC6, FT7, FT8) of the extended

10 − 20 system were omitted to allow the HMD to be

mounted. Thus, 60 channels were used for the record-

ing.

The actual prediction system was active in four of

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

ten sessions (online runs) and inactive in six out of ten

sessions. Thus, the overall data consisted of 4 online

and 6 pseudo-online sessions. Pseudo-online sessions

consisted of three, and online sessions of four runs,

with the forth run being the actual online run. Other-

wise they were treated in the same way. A single run

lasted in average about 13.64 ± 3.85 min.

The data acquired in the last run of each session

was used for evaluation of our system, and the other

sessions for training of the system. The pseudo-online

sessions were analyzed ofﬂine after the experiments

and are used here to provide a more comprehensive

data basis.

The runs contained between 466 and 1553 (in av-

erage ≈ 865) standards, between 16 and 51 (in aver-

age ≈ 35) targets, and between 1 and 54 (in average

≈ 8) missed targets.

All processing was performed on equally-shaped

windows of data with 1 s of duration, which were cut-

out and labeled according to the occurrence of a stan-

dard, target or missed target.

Between two runs there was a short break of

2 − 3 min, except before the last run in the online ses-

sions where the movement prediction system had to

be trained and thus the break lasted around 10 min.

3.2 Processing Methods

We used our software pySPACE (Signal Processing

And Classiﬁcation Environment) (Krell et al., 2013a;

Krell et al., 2013b)

for online and pseudo-online

data analysis.

All windows were processed independently from

each other. The data was preprocessed in several steps

in order to extract the relevant features for the clas-

siﬁer. First, the data was standardized channel-wise

by subtracting the mean signal value of the channel

and divided by the standard deviation of the channel

in the corresponding signal window. Next, a decima-

tion with an anti-alias ﬁnite impulse response ﬁlter

was performed to reduce the sampling rate of the data

from 5 kHz to 25 Hz. This was followed by another

band pass ﬁlter with pass band from 0.1 to 4.0 Hz.

Afterwards, the dimension of the data was reduced

further in several steps. First, the xDAWN spatial ﬁl-

ter (Rivet et al., 2009) was applied to reduce the 60

channels data to eight channels. The xDAWN spatial

ﬁlter splits the data in noise and signal-plus-noise sub-

spaces to, at the same time, extract meaningful sig-

nal information and reduce the dimension of the data.

The number of eight channels was chosen based on

previous experience with other experimental setups.

available at http://pyspace.github.com/pyspace

We used straight lines, a special form of local

polynomial features, as features for the classiﬁer. The

straight lines were ﬁt channel-wise to segments of the

data. Each segment lasted for about 400

ms and adja-

cent segments overlapped by about 400

ms. The slope

of the lines were used as features. These were stan-

dardized again in the next processing step.

We used a soft-margin support vector ma-

chine (SVM) with a linear kernel for classiﬁcation,

where the complexity regularization hyperparameter

was optimized using a grid search (tested values:

−6

, 10

−5

, . . . , 10

) and an internal 5-fold cross val-

idation. The data acquired in the training runs (see

Sec. 3.1) were used for the training and hyperparam-

eter optimization.

3.3 Evaluation

The data was evaluated with respect to the ability of

a classiﬁer to distinguish online and in single trial be-

tween EEG examples that contain patterns related to

target recognition processes (that evoke a P300) and

EEG examples that miss these patterns.

We expect that both, missed targets and standards

would not evoke such patterns while targets do. Since

enough training data was available we ﬁrst analyzed

how well a classiﬁer trained on standard and tar-

get examples performs to distinguish between both

classes (TS case).

Afterwards, we used a classiﬁer that was trained

on standard and target examples to distinguish be-

tween target and missed target examples (TM case)

which was the relevant application case (see Sec. 2).

To compare both results allows us to estimate how

well classiﬁer transfer performs in a demanding ap-

plication teleoperation scenario.

4 RESULTS AND DISCUSSION

For the evaluation of the classiﬁcation performance

we use the Balanced Accuracy (BA), which is given

T PR +

T NR, where T PR and T NR are the true

positive and true negative rate, respectively. Accord-

ingly, a BA score of 0.5 corresponds to random guess-

ing, while a BA score of 1.0 would correspond to a

perfect classiﬁer. The BA is not affected by an unbal-

anced number of examples in each class, as it is the

case here.

The obtained classiﬁcation performance (as BA)

for the different runs and sessions is shown in Table

1. The results correspond to the single trial detection

of the P300 ERP.

OnlineDetectionofP300relatedTargetRecognitionProcessesDuringaDemandingTeleoperationTask-Classifier

TransferfortheDetectionofMissedTargets

Table 1: Classiﬁcation performance for target vs. standard (TS case) and target vs. missed targets (TM case). The ﬁrst 6

columns on the left contain the results for the pseudo-online sessions, the 4 columns on the right contain the results for the

online sessions.

Evaluation S1 R1 S1 R2 S2 R1 S2 R2 S3 R1 S3 R2 S1 R

1 S1 R

2 S1 R

3 S2 R

TS 0.84 0.97 0.96 0.95 0.88 0.86 0.91 0.88 0.85 0.93

TM 0.64 0.98 0.86 0.72 0.83 0.78 0.81 0.9 0.80 0.94

The average classiﬁcation performance (as BA)

in the TM case is 0.827 ± 0.103, which is slightly

smaller as in the TS case (0.902 ± 0.048). It can be

observed that in 8 of 10 sessions the difference of the

BA is 0.1 or less, and in 5 of 10 sessions it is even

0.05 or less.

This shows that the achieved classiﬁcation perfor-

mance of our system works well for the P300 single

trial detection for both the TM and TS cases despite

the disturbance-prone setup. In addition, there is no

remarkable difference between the TM and TS cases.

Furthermore, the difference between the BA score

in the pseudo-online and online sessions is negligible

(average classiﬁcation performance TS case pseudo-

online: ≈ 0.9089 ± 0.057, online: ≈ 0.893 ± 0.037;

TM case pseudo-online: ≈ 0.803 ± 0.110, online ≈

0.864 ± 0.070).

5 CONCLUSIONS AND FUTURE

WORK

The presented results show that it is possible to de-

tect the P300 in a complex and noisy application sce-

nario where the operator of a robot has to perform a

dual task, i.e., to teleoperate a robot and to respond to

warnings. Furthermore, our results show that a clas-

siﬁer can be transferred between classes in case that

both classes, here standard and missed targets, that

miss a prominent pattern in the EEG signal, here the

P300 ERP. This transfer works for most cases very

well without a decrease of classiﬁcation performance.

However, in some cases the performance does de-

crease by a larger amount. Causes for this have to

be investigated.

In future, we plan to improve the signal process-

ing and pattern recognition methodology further to re-

duce the amount of required training data and to com-

pensate for changes of the user itself, like fatigue, by

applying online adaptation methods. In addition, we

plan to use and evaluate our system in even more ad-

vanced and complex application scenarios, e.g., the

supervision of operators that control several robots si-

multaneously.

ACKNOWLEDGEMENTS

This work was funded by the Federal Ministry of Eco-

nomics and Technology (BMWi, grant no. 50 RA

1012 and 50 RA 1011).

REFERENCES

Farwell, L. A. and Donchin, E. (1988). Talking off the

top of your head: toward a mental prosthesis utiliz-

ing event-related brain potentials. Electroencephalogr.

Clin. Neurophysiol., 70(6):510–23.

Folgheraiter, M., Jordan, M., Straube, S., Seeland, A., Kim,

S.-K., and Kirchner, E. A. (2012). Measuring the im-

provement of the interaction comfort of a wearable ex-

oskeleton. International Journal of Social Robotics,

4(3):285–302.

George, L. and L

ecuyer, A. (2010). An overview of research

on ”passive” brain-computer interfaces for implicit

human-computer interaction. In International Con-

ference on Applied Bionics and Biomechanics ICABB

2010 - Workshop W1 ”Brain-Computer Interfacing

and Virtual Reality”, Venice, Italy.

Guger, C., Harkam, W., Hertnaes, C., and Pfurtscheller, G.

(1999). Prosthetic control by an EEG-based brain-

computer interface (BCI). 5th European AAATE Con-

ference.

Haufe, S., Treder, M. S., Gugler, M. F., Sagebaum, M., Cu-

rio, G., and Blankertz, B. (2011). EEG potentials pre-

dict upcoming emergency brakings during simulated

driving. Journal of Neural Engineering, 8(5):056001.

Iturrate, I., Montesano, L., and Minguez, J. (2013). Task-

dependent signal variations in EEG error-related po-

tentials for brain-computer interfaces. Journal of Neu-

ral Engineering, 10:026024.

Kim, S. K. and Kirchner, E. A. (2013). Classiﬁer transfer-

ability in the detection of error related potentials from

observation to interaction. In Proceedings of the IEEE

International Conference on Systems, Man, and Cy-

bernetics, SMC 2013, Manchester, UK, October 13-

16, 2013.

Kirchner, E. A. and Drechsler, R. (2013). A Formal Model

for Embedded Brain Reading. volume 40, pages 530–

540. Emerald Group Publishing Limited.

Kirchner, E. A. and Kim, S. K. (2012). EEG in Dual-Task

Human-Machine Interaction: Target Recognition and

Prospective Memory. In Proceedings of the 18th An-

nual Meeting of the Organization for Human Brain

Mapping.

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

Kirchner, E. A., Kim, S. K., Straube, S., Seeland, A.,

ohrle, H., Krell, M. M., Tabie, M., and Fahle, M.

(accepted 2013). On the applicability of brain read-

ing for self-controlled, predictive human-machine in-

terfaces in robotics. PLoS ONE.

Kirchner, E. A., Metzen, J. H., Duchrow, T., Kim, S. K.,

and Kirchner, F. (2009). Assisting Telemanipulation

Operators via Real-Time Brain Reading. In Lohweg,

V. and Niggemann, O., editors, Proc. Mach. Learn-

ing in Real-time Applicat. Workshop 2009, Lemgoer

Schriftenreihe zur industriellen Informationstechnik,

Paderborn, Germany.

Kirchner, E. A., W

ohrle, H., Bergatt, C., Kim, S.-K., Met-

zen, J. H., Feess, D., and Kirchner, F. (2010). Towards

operator monitoring via brain reading – an EEG-based

approach for space applications. In Proc. 10th Int.

Symp. Artiﬁcial Intelligence, Robotics and Automa-

tion in Space, pages 448–455, Sapporo.

Krell, M. M., Straube, S., Seeland, A., W

ohrle, H., Teiwes,

J., Metzen, J. H., Kirchner, E. A., and Kirchner, F.

(2013a). pySPACE. https://github.com/pyspace.

Krell, M. M., Straube, S., Seeland, A., W

ohrle, H., Tei-

wes, J., Metzen, J. H., Kirchner, E. A., and Kirchner,

F. (submitted 2013b). pySPACE a signal processing

and classiﬁcation environment in Python. Frontiers in

Neuroinformatics.

Kutas, M., McCarthy, G., and Donchin, E. (1977). Aug-

menting mental chronometry: the P300 as a measure

of stimulus evaluation time. Science, 197(4305):792–

Nijholt, A., Tan, D., Allison, B. Z., Del R Milan, J., and

Graimann, B. (2008). Brain-computer interfaces for

HCI and games, pages 3925–3928. ACM.

Pan, S. J. and Yang, Q. (2010). A Survey on Transfer Learn-

ing. IEEE Transactions on Knowledge and Data En-

gineering, 22(10):1345–1359.

Polich, J. (2007). Updating P300: an integrative theory of

P3a and P3b. Clin Neurophysiol, 118(10):2128–48.

Reuderink, B. (2008). Games and Brain-Computer Inter-

faces: The State of the Art. WP2 BrainGain Deliver-

able HMI University of Twente September 2008, pages

1–11.

Rivet, B., Souloumiac, A., Attina, V., and Gibert, G. (2009).

xDAWN algorithm to enhance evoked potentials: ap-

plication to brain-computer interface. IEEE Transac-

tions on Biomedical Engineering, 56(8):2035–2043.

Salisbury, D., Rutherford, B., Shenton, M., and McCar-

ley, R. (2001). Button-pressing Affects P300 Am-

plitude and Scalp Topography. Clin Neurophysiol,

112(9):1676–1684.

Seeland, A., W

ohrle, H., Straube, S., and Kirchner, E. A.

(2013). Online movement prediction in a robotic

application scenario. In Neural Engineering (NER),

2013 6th International IEEE/EMBS Conference on.

Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,

Pfurtscheller, G., and Vaughan, T. M. (2002). Brain-

computer interfaces for communication and control.

Clin. Neurophysiol., 113(6):767–91.

Zander, T. O., Kothe, C., Jatzev, S., and Gaertner, M.

(2010). Enhancing human-computer interaction with

input from active and passive brain-computer inter-

faces. Brain-Computer Interfaces, pages 181–199.

OnlineDetectionofP300relatedTargetRecognitionProcessesDuringaDemandingTeleoperationTask-Classifier

TransferfortheDetectionofMissedTargets