Content uploaded by L. Abril Torres-Méndez
Author content
All content in this area was uploaded by L. Abril Torres-Méndez on Mar 26, 2015
Content may be subject to copyright.
Are You Talking to Me?
Detecting Attention in First-Person Interactions
Luis C. Gonz´
alez-Garc´
ıa
and L. Abril Torres-M´
endez
Robotics and Advanced Manufacturing Group
CINVESTAV Campus Saltillo
Ramos Arizpe, M´
exico
Email: carlos.gonzalez@cinvestav.edu.mx
abril.torres@cinvestav.edu.mx
Julieta Martinez, Junaed Sattar
and James J. Little
Department of Computer Science
The University of British Columbia
Vancouver, Canada
Email: {julm,junaed,little}@cs.ubc.ca
Abstract—This paper presents an approach for a mobile robot
to detect the level of attention of a human in first-person
interactions. Determining the degree of attention is an essential
task in day-to-day interactions. In particular, we are interested
in natural Human-Robot Interactions (HRI’s) during which a
robot needs to estimate the focus and the degree of the user’s
attention to determine the most appropriate moment to initiate,
continue and terminate an interaction. Our approach is novel in
that it uses a linear regression technique to classify raw depth-
image data according to three levels of user attention on the
robot (null, partial and total). This is achieved by measuring
the linear independence of the input range data with respect to a
dataset of user poses. We overcome the problem of time overhead
that a large database can add to real-time Linear Regression
Classification (LRC) methods by including only the feature
vectors with the most relevant information. We demonstrate
the approach by presenting experimental data from human-
interaction studies with a PR2 robot. Results demonstrate our
attention classifier to be accurate and robust in detecting the
attention levels of human participants.
Keywords–Human-robot interaction; Body pose classification;
Least squares approximations; Raw range data analysis.
I. INTRODUCTION
Determining the attention of people is an essential compo-
nent of day-to-day interactions. We are constantly monitoring
other people’s gaze, head and body poses while engaged in a
conversation [1][2][3]. We also perform attention estimation in
order to perform natural interactions [4][5]. In short, attention
estimation is a fundamental component of effective social
interaction; therefore, for robots to be efficient social agents
it is necessary to provide them with reliable mechanisms to
estimate human attention.
We believe that human attention estimation, particularly
in the context of interactions, is highly subjective. However,
attempts to model it have been relatively successful, e.g.,
allowing a robot to ask for directions when it finds a human,
as in the work of Weiss et al. [6]. Nonetheless, the state-
of-the-art is still far from reaching a point where a robot
can successfully interact with humans without relying on
mechanisms not common to natural language. Recently, the
use of range images to make more natural human-machine
interfaces has been in the agenda of researchers, like in the
case of the Microsoft KinectTM, which delivers a skeleton of
Figure 1. Left: Raw range input that a robot gets when trying to asses
human attention, as described in this work. Right: Set-up scenario for our
experiments. The PR2 robot approaches a human sitting at a desk..
a human that can be further used as a high-level feature of the
human pose [7]. Although good results have been obtained
with such devices in pose estimation, little effort has been
devoted to further infer information about the user from such
data. In this work, we use range data (similar to that shown
in Figure 1) to infer the level of attention of the user, which
is not explicitly given by the sensor output.
Our approach is novel in that it uses raw depth images to
evaluate the attention level of a subject, regardless of whether
she is facing the depth sensor, in order to classify her pose in
an attention scale. In this work, we focus on learning human
attention from raw depth images by using the LRC algorithm,
which can be exploited by social robots to determine the best
moment to ask for support from a human sitting at her desk,
like those found in common working or reading spaces.
The remainder of this paper is structured as follows: In
Section II, we talk about how other authors have tackled the
problem of attention awareness detection using images and
range information as a source. In Section III, we describe
the problem that this paper faces, attention estimation from
a first person perspective using only raw range information.
In Section IV, we walk through the technical aspects of
the methodology that we propose (LRC). In section V, it is
described the actual set-up and execution of the experiments,
as well as the interpretation and discussion of the data gath-
ered from them. Finally, in Section VI, our conclusions and
suggestions for future work are exposed.
II. RE LATE D WOR K
The problem of attention awareness detection, despite its
relevance, remains largely unexplored in the HRI literature.
Here we present some of the building blocks of our work.
137Copyright (c) IARIA, 2015. ISBN: 978-1-61208-390-2
COGNITIVE 2015 : The Seventh International Conference on Advanced Cognitive Technologies and Applications
A. Pose, Head and Gaze Estimation
Some of the most effective social cues for attention estima-
tion are gaze, body and head poses. Fortunately, a large body
of knowledge has been gathered in these areas.
Shotton et al. [7], used a single Red-Green-Blue+Depth
(RGB+D) camera to perform pose estimation and body parts
recognition. A randomized decision forest was trained on
synthetic data that covered a wide range of human poses and
shapes. Features were obtained by computing the difference
of depth between two points. A further speedup was achieved
by providing a GPU implementation. Vision-only approaches
range from the Flexible Mixtures-of-Parts [8], an extension
of the Deformable Parts Model [9] which explicitly accounts
for different body deformations and appearances, to leverage
poselet-based part detections for further constraining optical-
flow-based-tracking of body parts [10].
Head pose estimation can be seen as a sub-field of full-
body pose estimation. In fact, Kondori, Yousefi, Haibo and
Sonning [11] extended the work of Shotton et al. to Head Pose
Estimation in a relatively straightforward manner. Similarly,
the problem becomes much harder when depth information is
no longer available.
For a more in-depth treatment of the subject, as well as
for recent advances in gaze estimation, we direct the reader to
the reviews made by Murphy-Chutorian and Trivedi [12] and
Hansen and Qiang [13].
B. Awareness Detection in Computer Vision
Estimating attention from visual input has been studied
particularly in the context of driving. Doshi and Trivedi [14]
built a system that incorporated cameras observing both the
human subject and her field of view. By estimating the gaze of
the subject and the saliency map from her viewpoint, they used
Bayes’ rule to obtain a posterior distribution of the location of
the subject’s attention. Our work is different from theirs since
just as in person-to-person interactions, we do not have access
to the field of view of the person, but we might rather be a
part of it.
Also related are Mutual Awareness Events (MAWEs).
MAWEs are events that concentrate the attention of a large
number of people at the same time. In this context, Benfold
and Reid [15] built upon evidence from the estimated head
poses of large crowds to guide a visual surveillance system
towards interesting points.
C. First-Person Interaction
Recently, some work has been devoted to transfer knowl-
edge gained from third to first person perspectives. Ryoo
and Matthies [16] performed activity recognition from a first-
person viewpoint from continuous video inputs. They com-
bined dense optical flow as a global descriptor and cuboids [17]
as local interest point detectors, then built a visual dictionary
to train an SVM classifier using multi-channel kernels.
D. Human Attention and Awareness Estimation
To this day, human attention remains an active area of
research. A widely accepted model of attention was proposed
by Itti and Koch [18], where attention is understood as the
mixture of “bottom-up”, i.e., unconscious, low-level features of
an image, and “top-down”, i.e., task-oriented mechanisms that
the subject controls consciously. Later work by Itti and Baldi
incorporated the element of Bayesian surprise [19], i.e., which
states things that are different on the temporal domain attract
attention, but with time they get incorporated into our world
model and become less relevant. We keep this factor in mind
when designing the experiment, as people who are not used to
interacting with a robot might direct their attention to it just
because it is something new, rather than because of its actions.
It is also important to mention that in order to attract the
user’s attention, the robot has to be attentive to the person. This
often involves mimicking human mechanisms that indicate
attention. Bruce, Nourbakhsh and Simmons [4] found that if
a robot turns its head to the person whose attention it wants,
then the probability of the person cooperating with the robot is
greatly increased. This is also exploited by Embgen et al. [20],
who further concluded that the robot needs only move its head
to transmit its emotional state. Nevertheless, no further analysis
was performed to determine whether or how this robot-to-
human non-verbal communication impacts HRI.
E. Linear Regression Classification
LRC is a simple yet powerful method for classification
based on linear regression techniques. Naseem, Togneri and
Bennamoun [21] introduced LRC to solve the problem of
face identification by representing an image probe as a linear
combination of class-specific image galleries. This is per-
formed by determining the nearest subspace classification and
solving the inverse problem to build a reconstructed image,
choosing the class with the minimum reconstruction error.
During the training phase, the inputs are added to the database
using a greedy approach. Every input image is required to
add a minimum information gain in terms of linear subspace
independence: only if they fulfill the criterion, they are added
to the database. This keeps the database size small, and allows
for efficient training and classification. To the best of our
knowledge, we are the first to apply this method to raw depth
image data.
III. PROB LE M FOR MU LATI ON
The problem that we address is human attention estimation
from a first-person perspective. At a coarse level, we define
attention in three categories: a) null attention, b) partial
attention and c) total attention. We believe that this simple
scale is enough to model a wide range of situations, since
they encode the willingness of a user to engage in interaction.
If robots are meant to be efficient social agents, it is imperative
to be able to detect the right instance to start, maintain and
end task-oriented interactions with humans.
A. Scenario
In our study, we assume that the robot wants to interact
with a human who is sitting at a desk, a common occurrence in
an office environment (see Figure 1); the robot wants to start an
interaction approaching the human from one side. The situation
is analog to a human approaching a coworker at her desk,
willing to know if she is available for a given task. For our
experiments, we use the Willlow Garage PR2 robot, with users
occupying lab workstations with computer terminals. Having
the experiment occur within the confines of the lab spaces
ensures users’ attentions are not unduly attracted to the robot,
as having the robot around is a fairly common occurrence in
the lab.
138Copyright (c) IARIA, 2015. ISBN: 978-1-61208-390-2
COGNITIVE 2015 : The Seventh International Conference on Advanced Cognitive Technologies and Applications
Figure 2. Representative raw-depth images of the three levels of attention. From left to right, null attention, partial attention and full attention.
The images were captured using a KinectTM mounted on the PR2 head.
B. Data
In order to evaluate human attention and obtain the best
moment to ask for support, the robot relies only on a set of
depth images captured with its KinectTM. For initial tests, the
data was captured using a separate KinectTM sensor mounted
on a tripod simulating the pose that the sensor would have
above the PR2. For our study, the data captured consisted
only of depth information, allowing our approach to be robust
against illumination variations, as well as other appearance
changes. The intention is to demonstrate that our approach
is robust enough so that visual information is not required,
and efficient enough to run on a constrained computational
platform while performing in real-time.
To build the training database, a subject is seated at a desk
and asked to perform activities that simulate the three levels of
attention of our scale. We describe each attention levels next:
1) Null attention (class 1): the subject’s posture is such
that she is facing the computer monitor, pretending
she is busy, working;
2) Partial attention (class 2): the subject’s posture is
such that she is not facing the monitor, nor the robot,
but rather facing somewhere in between;
3) Full attention (class 3): the subject’s posture is such
that she is facing the robot.
The subject is free to simulate the three attention levels
according to her discretion, as long as the basic guidelines
described above are satisfied. We recorded the movements of
the subject and repeated the experiment several times; each
one by placing the depth sensor in different configurations
(i.e., changing position, elevation and orientation), allowing
for a more versatile training set. Examples of the range data
are shown in Figure 2.
IV. TECHNICAL APP ROACH
Our approach towards determining attention levels consists
of an offline training stage and an online detection stage.
The training step includes capturing depth snapshots (or video
streams) of the user at her workstation or desk, extracting
features and constructing class-specific feature matrices to
build an attention classifier. During attention classification,
instantaneous depth image snapshots from the KinectTM are
fed into the classifier, and the class with the minimum linear
independence with respect to the training data is chosen as the
likely attention level. The following sections provide technical
details of these individual steps.
Figure 3. Left: Raw depth image and Right: preprocessed image
(γ= 1/20 and distance = 2meters).
A. Interaction Setting
The HRI is carried out as follows. First, the robot ap-
proaches and stands on either side of the human, using the
range data to assess if the human is occupied, and if that
is the case, continue evaluating the best moment to ask for
support. If the human remains busy for an extended duration,
then the robot does not engage in interaction and attends to
other tasks. The aim is to evaluate if a robot standing close to
a human working at a desk can accurately estimate the degree
of attention of the human, and use this information to ask for
support at the correct time instance. Data for our training set
is obtained from the KinectTM mounted on the robot in such
scenarios, and is limited to range data only. Collected range
images consist of particpants performing actions corresponding
to the attention levels that we defined. The image sensor is
placed on both sides of the user (see Figure 4), while recording
the actions of the subject.
B. Features
For our LRC, the features consist of depth image data
downsampled to γ=1
20 scale (see Figure 3), and reshaped
by concatenation of its columns, similar to the methodology
of Naseem, Togneri and Bennamoun [21]. However, as we are
working with depth images, and we do not want the scene
background to interfere with the learning and classification
processes, the depth images are preprocessed to remove un-
wanted data. Specifically, we consider only those depth values
up to a specific range, which accounts for the approximate
physical distance between a robot and a human sitting at his
desk under the current interaction scenario. This distance was
empirically observed to be approximately 2meters from the
KinectTM sensor.
139Copyright (c) IARIA, 2015. ISBN: 978-1-61208-390-2
COGNITIVE 2015 : The Seventh International Conference on Advanced Cognitive Technologies and Applications
Figure 4. Range data from the left and right profile of the subjects.
Both were included on the database for this study.
C. Training
Naseem, Togneri and Bennamoun [21] consider a database
with photographs of people’s faces. This is translated into a
small number of sample images per class (subject), avoiding
the time constraint of solving the pseudoinverse involved on
a linear regression, experienced on large datasets, like videos.
The novelty of our work is that we deal with this constraint (big
datasets) by analyzing the linear independence of the images.
By doing this, we dismiss all the new pictures that does not
add relevant information to the LRC. Thus, we can condense
the dataset into representative images, without losing relevant
information. This allows us in turn to achieve an efficient LRC
in real time.
Let Y ∈Rmbe a vector of a given matrix X ∈Rm×n.
We used a linear regression technique to analyze the linear
independence of Y, as shown in Algorithm 3. In this algorithm,
vector Y is projected onto the column space of X, then an
error is calculated by subtracting the resultant projected vector
Ycto the original vector Y, giving as result the projection
error ε. This εis a metric that is used to measure the linear
independence of Y with respect to the column space of X.
D. Algorithm
The overall algorithm is divided into two parts,
1) Build X: This procedure (Algorithm 1) analyzes the
range image database and builds one matrix per atten-
tion class that contains the most significant collection
of images in that class, and ensures a maximum
degree of linear independence between images in the
same class.
2) Classify Y: This procedure (Algorithm 4) is respon-
sible for using the class matrices generated by the
Build X algorithm, as well as the input depth image,
to classify that image into one of the known classes.
It also outputs the projection error of the image with
respect to the column space generated by each of the
class matrices, choosing the class with the minimum
projection error.
It is important to mention that in order to reduce the
overhead of a pseudo inverse calculation, X†
iis calculated only
when Xichanges, thus Xiand X†
iare saved and computed only
once for classification.
V. EX PE RI ME NTAL RESULTS
We conducted a number of trials to evaluate our proposed
approach. To train our system, we used range data of the three
specific attention classes from 5different participants, follow-
ing the process described in Section IV. For each participant
in the training process, we obtained video streams for three
different attention levels, in two different KinectTM positions;
Algorithm 1 Build X, the probe database.
Require: Threshold τ
1: for each class ido
2: Img ←Random unseen image. .Initialize Xi
3: Y←FE ATUR ES(Img )
4: Append Y to Xi
5: Compute X†
i.Build Xi
6: for each new image nImage do
7: Y←FE ATUR ES(nImage)
8: ε←LINEARINDEPENDENCE(Y,Xi,X†
i)
9: if ε≥τthen
10: Append Y to Xi
11: Compute X†
i
12: end if
13: end for
14: end for
15: save(X†
i, X)
Algorithm 2 Feature extraction. Performs downsampling and
reshaping.
1: procedure FEATU RE S( Image, γ)
2: Cut the image background.
3: Down-sample the image by a factor γ.
4: Reshape the image to a column vector.
5: return The post-processed Image.
6: end procedure
each of the video streams have dimensions of 640×480 pixels,
and have approximately 500 frames each. This resulted in a
total of 15,000 frames for the training process. The attention
matrices Xihave average dimensions of 39 ×768, with 39
images downscaled to 1
20 of their original dimensions. The
value of the threshold τwas empirically set at 4.0for all
trials. Training and classification was performed on a PC with
an Intel Core-i5TM processor running at 1.7GHz, with 2GB
of memory and under the Ubuntu 12.04 Long-term Release
(LTS) edition. The code was implemented in C++ using the
Robot Operating System (ROS) C++ bindings.
Our results are summarized in Figures 5 and 6. The figures
show frame-by-frame reconstruction errors of a test video.
The errors represents the linear independence of an input
image with respect to the column space of each class-specific
database Xi,i∈ {1,2,3}.
Figure 5(a) shows the classification of a video that was
used during the training phase, as expected, the projection
error is close to zero almost all the time. When the error
reaches zero, is an indication that the current image passed
the linear independence test, and it was used on the learning
phase. While this is not illustrative of the accuracy of our
algorithm, it does illustrate the fact that most of the information
used during training is redundant, and that by keeping only
a small fraction of it we can achieve a low reconstruction
error. Figure 5(b) shows the classification performance on a
video sequence that was not used during the training phase.
While the reconstruction error is larger in this case, it is
nonetheless sufficient to perform classification accurately. The
key observation is that irrespective of the actual reprojection
error numbers, there is clear separation between the actual
attention class errors and the errors of the other attention
classes, which leads to distinct identification of the user’s
attention level.
140Copyright (c) IARIA, 2015. ISBN: 978-1-61208-390-2
COGNITIVE 2015 : The Seventh International Conference on Advanced Cognitive Technologies and Applications
Algorithm 3 Measure the linear independence of Y with
respect to database X.
Require: Y∈Rm, X, ∈Rm×n, X†∈Rn×m
1: procedure LINEARINDEPENDENCE(Y, X, X†)
2: Yc←XX†Y.Reprojection of Y.
3: ε← ||Yc−Y||2
4: return ε . The reprojection error is the metric.
5: end procedure
Algorithm 4 Classify a new vector Y.
Require: An input Image. Precomputed Xiand X†
ifor each class i,
the class-specific databases and their pseudo-inverses.
1: for each class ido
2: Y←FE ATUR ES( Image )
3: εi←LI NEARINDEPENDENCE(Y,Xi,X†
i)
4: end for
5: return argmini(εi).Return the class with minimum error.
The LRC is done per frame, by choosing the minimum
of these errors, and using a leave-one-out validation process
during training (i.e., while building X, one subject was left
out the matrix). Hence, as is observed in Figure 5, the robot is
capable of correctly estimating the attention level, even when
testing on a subject that was not included on the database.
Figure 7 shows the difference in poses between the training
(left) and testing (right) sets.
Nevertheless, large variations in the RGB+D sensor pose
can lead to reduced performance of our algorithm. This is
demonstrated in Figure 6, where the robot (and thus the
KinectTM) is continually placed in positions not used to capture
training data, while the system tries to detect a user in Class 3
(i.e., full) attention level. As the KinectTM changes its pose,
the errors levels vary, resulting in an inaccurate classification.
Note that the seperations between classes on the error scale
are also reduced, resulting in degraded accuracy. The slopes
on the figure represent displacement of the KinectTM.
A. Quantitative Analysis
In order to evaluate and validate our proposed method, we
compare it against other common approaches for estimating
visual attention based only on visual information, namely the
Head Pose and Gaze attention estimation [22][14]. Ideally
the comparison should be carried out using a ground truth
of the subject attention, but this is extremely subjective, due
to the inherently complexity of the human behavior, for this
reason a simulated labeled attention is used as ground truth.
To simulate this baseline, attention is lurked to the camera by
showing interesting images to the user in a display beneath
the sensor; similarly the attention is also directed outside the
camera showing interesting images in another display.
For the purpose of this evaluation, the three attentive states
are wrapped into two main states, attention or no attention
towards the sensor, so, when the estimator and the simulated
ground truth coincide on the attentive state of the user, a
1or OK is recorded, otherwise a 0or NOT OK is recorded.
In the end, all this records are averaged in order to obtain an
average accuracy of the estimator against the simulated ground
truth, which in the context of this paper, it can be used as the
estimator’s main metric of performance.
The performance comparison is summarized on the Table I.
(a) Included in the dataset
(b) Unkown subject
Figure 5. Reprojection class error vs. Time. An input video from Class 1
(no attention) is being classified in real time. (a) LRC over a dataset that
was included on the learning database. (b) LRC over data that was not
included on the learning database.
TABLE I. Comparison of proposed Raw-Range-Information attention
estimator with approaches based purely on visual information, Head Pose
and Eye Gaze (coarse). Avg. Accuracy corresponds to the attention state
between that reported by the estimator and the labeled attention.
Estimator Avg. Accuracy
HeadPose-based 0.7333
EyeGaze-based 0.5667
LRC on raw range (proposed) 0.8500
VI. CONCLUSIONS AND FUTURE WO RK
We have presented a novel method to accurately estimate
the attention of a person when interacting with a robot.
The results demonstrate that our method outperforms other
models for attention estimation based solely on visual infor-
mation (Table I).
The estimation is performed using raw depth data of the
person’s body pose. We use a LRC with the selection of
the best linearly independent depth-images during the train-
ing phase. Our method also effectively discards redundant
information during the training phase, while maintaining good
performance on previously-unseen sequences. In addition to
being completely independent from appearance and illumina-
tion changes, our approach is robust to small pose variations,
141Copyright (c) IARIA, 2015. ISBN: 978-1-61208-390-2
COGNITIVE 2015 : The Seventh International Conference on Advanced Cognitive Technologies and Applications
Figure 6. Effect of sensor displacement on classifier performance
(Full attention).
Figure 7. Difference in the position of the range sensor between images that
were included (left) and not included (right) on the database for this study.
from the training data. The main advantages of our classifier
are its simplicity, real-time computability and small memory
footprint, which is ideal for implementation on board robots
for man-machine interaction tasks.
In future work, we intend to explore attention estimation
in a wider variety of settings, perform larger-scale experiments
(encompassing more attention classes, more subjects, more
situations), and explore the limits of the LRC approach.
Particularly, we plan to fuse this approach with other face
and gaze detectors in order to achieve a short-long distance
attention estimator.
ACK NOW LE DG EM EN TS
We would like to thank CONACyT and CBIE, under the
ELAP, as well as Fonds de Recherche du Queb´
ec – Nature et
Technologie, for their support and project funding.
REFERENCES
[1] M. Argyle, Bodily communication. Routledge, 2013.
[2] M. Knapp, J. Hall, and T. Horgan, Nonverbal communication in human
interaction. Cengage Learning, 2013.
[3] N. Hadjikhani, K. Kveraga, P. Naik, and S. P. Ahlfors, “Early (n170)
activation of face-specific cortex by face-like objects,” Neuroreport,
vol. 20, no. 4, 2009, p. 403.
[4] A. Bruce, I. Nourbakhsh, and R. Simmons, “The role of expressiveness
and attention in human-robot interaction,” in Robotics and Automation,
2002. Proceedings. ICRA ’02. IEEE International Conference on, vol. 4,
2002, pp. 4138–4142.
[5] S. Lang, M. Kleinehagenbrock, S. Hohenner, J. Fritsch, G. A. Fink,
and G. Sagerer, “Providing the basis for human-robot-interaction: A
multi-modal attention system for a mobile robot,” in Proceedings
of the 5th International Conference on Multimodal Interfaces, ser.
ICMI ’03. New York, NY, USA: ACM, 2003, pp. 28–35. [Online].
Available: http://doi.acm.org/10.1145/958432.958441
[6] A. Weiss, J. Igelsbock, M. Tscheligi, A. Bauer, K. Kuhnlenz, D. Woll-
herr, and M. Buss, “Robots asking for directions - the willingness of
passers-by to support robots,” in Human-Robot Interaction (HRI), 2010
5th ACM/IEEE International Conference on, March 2010, pp. 23–30.
[7] J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore,
A. Kipman, and A. Blake, “Real-time human pose recognition in parts
from single depth images,” in Computer Vision and Pattern Recognition
(CVPR), 2011 IEEE Conference on, June 2011, pp. 1297–1304.
[8] Y. Yang and D. Ramanan, “Articulated pose estimation with flexi-
ble mixtures-of-parts,” in Computer Vision and Pattern Recognition
(CVPR), 2011 IEEE Conference on, June 2011, pp. 1385–1392.
[9] P. Felzenszwalb, D. McAllester, and D. Ramanan, “A discriminatively
trained, multiscale, deformable part model,” in Computer Vision and
Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, June
2008, pp. 1–8.
[10] K. Fragkiadaki, H. Hu, and J. Shi, “Pose from flow and flow from
pose,” in Computer Vision and Pattern Recognition (CVPR), 2013 IEEE
Conference on, June 2013, pp. 2059–2066.
[11] F. Kondori, S. Yousefi, H. Li, and S. Sonning, “3d head pose estimation
using the kinect,” in Wireless Communications and Signal Processing
(WCSP), 2011 International Conference on, Nov 2011, pp. 1–4.
[12] E. Murphy-Chutorian and M. Trivedi, “Head pose estimation in com-
puter vision: A survey,” Pattern Analysis and Machine Intelligence,
IEEE Transactions on, vol. 31, no. 4, April 2009, pp. 607–626.
[13] D. Hansen and Q. Ji, “In the eye of the beholder: A survey of models
for eyes and gaze,” Pattern Analysis and Machine Intelligence, IEEE
Transactions on, vol. 32, no. 3, March 2010, pp. 478–500.
[14] A. Doshi and M. Trivedi, “Attention estimation by simultaneous obser-
vation of viewer and view,” in Computer Vision and Pattern Recognition
Workshops (CVPRW), 2010 IEEE Computer Society Conference on,
June 2010, pp. 21–27.
[15] B. Benfold and I. Reid, “Guiding visual surveillance by tracking
human attention,” in Proceedings of the 20th British Machine Vision
Conference, September 2009, pp. 1–11.
[16] M. Ryoo and L. Matthies, “First-person activity recognition: What
are they doing to me?” in Computer Vision and Pattern Recognition
(CVPR), 2013 IEEE Conference on, June 2013, pp. 2730–2737.
[17] P. Dollar, V. Rabaud, G. Cottrell, and S. Belongie, “Behavior recog-
nition via sparse spatio-temporal features,” in Visual Surveillance and
Performance Evaluation of Tracking and Surveillance, 2005. 2nd Joint
IEEE International Workshop on, Oct 2005, pp. 65–72.
[18] L. Itti and C. Koch, “Computational modelling of visual attention,”
Nature reviews neuroscience, vol. 2, no. 3, 2001, pp. 194–203.
[19] L. Itti and P. Baldi, “Bayesian surprise attracts human attention,” Vision
research, vol. 49, no. 10, 2009, pp. 1295–1306.
[20] S. Embgen, M. Luber, C. Becker-Asano, M. Ragni, V. Evers, and
K. Arras, “Robot-specific social cues in emotional body language,” in
RO-MAN, 2012 IEEE, Sept 2012, pp. 1019–1025.
[21] I. Naseem, R. Togneri, and M. Bennamoun, “Linear regression for
face recognition,” Pattern Analysis and Machine Intelligence, IEEE
Transactions on, vol. 32, no. 11, Nov 2010, pp. 2106–2112.
[22] A. Doshi and M. Trivedi, “Head and gaze dynamics in visual attention
and context learning,” in Computer Vision and Pattern Recognition
Workshops, 2009. CVPR Workshops 2009. IEEE Computer Society
Conference on, June 2009, pp. 77–84.
142Copyright (c) IARIA, 2015. ISBN: 978-1-61208-390-2
COGNITIVE 2015 : The Seventh International Conference on Advanced Cognitive Technologies and Applications
