=Paper=
{{Paper
|id=None
|storemode=property
|title=Human Action Recognition Oriented to Humanoid Robots Action Reproduction
|pdfUrl=https://ceur-ws.org/Vol-860/paper12.pdf
|volume=Vol-860
|dblpUrl=https://dblp.org/rec/conf/aiia/MichielettoM12
}}
==Human Action Recognition Oriented to Humanoid Robots Action Reproduction==
https://ceur-ws.org/Vol-860/paper12.pdf
Human action recognition
oriented to humanoid robots action reproduction
Stefano Michieletto and Emanuele Menegatti
Intelligent Autonomous Systems Lab (IAS-Lab)
Department of Information Engineering (DEI)
Faculty of Engineering, The University of Padua
Via Ognissanti 72, I-35129 Padova, Italy
{stefano.michieletto,emg}@dei.unipd.it
http://robotics.dei.unipd.it
Abstract. Our research aims at providing a humanoid robot with the
ability of observing, learning, and reproducing actions performed by hu-
mans in order to acquire new skills. In other words, we want to apply
artificial intelligence techniques to automatically recognize a human ac-
tivity in order to make a humanoid robot able to reproduce it.This sys-
tem has not only to distinguish between different actions, but also to
represent them in a proper manner to allow a robot to reproduce the
motion trajectories the demonstrator showed and learn new skills. Since
the final system is going to be integrated in an autonomous humanoid
robot (specifically model Aldebran Nao), we are working with an RGB-D
sensor (Microsoft Kinect) that can be easily applied to it. This objec-
tive introduces also strict real-time constrains to the action recognition
algorithm: we have opted for a probabilistic approach that offers good
modeling and fast recognition performances.
Keywords: Action recognition, humanoid robots, programming by demon-
stration, imitation learning
1 Introduction
When the Turing Test was proposed, in 1950 [9], the idea was to compare humans
and machines by looking at their responses. We would like to extend this criteria
and compare humans and machines by looking at their actions. This regards the
abilities to perceive the world, to learn from what happen around themselves,
and the way they move with respect to the rest of the environment, that is
the hallmark of an intelligent system. This project aims to provide a robot the
capability to perform a task, recognizing actions from human demonstrators and
learning from their movements.
A lot of work has been done from the 1950 in these fields, in particular
in computer vision and robotics research. Computer vision goal is to extract
information from a scene and structure it, and action recognition is assuming
more and more importance in this field for its correlation with wide range of
application such as biometrics, animation and interactive environments, video
analysis and surveillance. The identification process can be divided into three
stages: feature extraction, feature representation and action classification [10].
Two different approaches characterize features extraction: model-based and with
no-model assumption. The first technique takes advantage of the prior knowl-
edge of the body model to extract the selected features from the scene. The
second one, instead, uses a common features extraction method. The features
representation is strictly connected to the extraction phase. The model-based
methods usually keep track of position and orientation of each part that com-
poses the model. No-model methods, on the other hand, maintain the features
characterize themselves: surface normals, RGB features, time-space information
of each point and so on. The action classification can be approached by compar-
ing static images with pre-stored samples (template-based [1]), by describing the
system with a series of states with a certain probability and allowing switches
between different states (probability-based [4][6]), or representing the motion with
a sequence of grammatical rules (grammar-based [2]).
Robotics was defined as the science which studies the intelligent connec-
tion between perception and action[8] and mixes together mechanics, controls,
computers and electronics, so that it benefits from advances in the different
technologies. Recently, robots has been used in more and more areas such as
service robotics, field robotics, or human augmentation. By the dawn of the new
millennium, the “human” factor pushes robots to rapidly assume anthropomor-
phic appearance and the more robots interact with people safer, smarter and
more independent they need to be, that why other disciplines were involved in
the robotics area. The human ability to imitate behaviors is already present in
even 12 - 21 day old infants [5]. Despite some hypothesis about the translation
between visual and motion features[11], the mechanisms that express a demon-
stration to a set of internal motor primitives remain largely unknown. Several
works in robotic imitation suggest that robots can perform new skills by learn-
ing key information from a demonstration [7]. More recently, this idea has been
applied on humanoid robots as a natural way to teach how to perform a task in
a dynamic environment [3].
The remainder of the paper is organized as follows: Section 2 describes the
acquisition system. In Section 3 the features extraction method and their repre-
sentation are introduced, while the action classification is described in Section
4. Conclusions and future works are contained in Section 5.
2 Acquisition system
The development of affordable RGB-D sensors is giving a rapid boost in com-
puter vision research. These sensors are also reliable: they allow to natively
capture RGB and depth information at good resolution and frame rate (the Mi-
crosoft Kinect takes 30 frames per second at 640x480 pixels of resolution). On
the other hand, the sunlight plays havoc with the infrared pattern projected by
the sensor and they cannot be used over eight meters. An indoor environment
is not a real limitation for the humanoid robot we will use, so mounting RGB-D
camera can be a very rich source of information for our work. The not invasive
impact is another advantage of such sensors with respect to motion sensors and
the robot can learn from every person in the scene.
In order to simplify the initial setting, the RGB-D sensor will not be immedi-
ately mounted on the robot. At first, the action recognition algorithm will work
as a stand-alone system, hence we can assume that the data can be elaborated
off-line. This assumption allows us to compare our algorithm with others at the
state of the art, without the constrain of a camera moving on a robot. For this
purpose, we are thinking of releasing a dataset of actions acquired as RGB-D
data. Figure 1 shows some people performing actions for our preliminary tests.
Once the algorithm will be efficient and robust enough, the robot movement will
be considered and the system will be adjusted to prevent the performances from
being affected by the robot movements.
(a) (b) (c) (d)
Fig. 1. Actions collected in our preliminary tests: pointing (a), sitting down(b), walking
(c), and waving (d)
3 Features extraction and representation
A model-based approach will be adopted for features extraction. At the beginning
of the project the OpenNI1 freeware implementation of skeleton tracker will
give us information about 25 skeletal joints covering the entire body. For each
joint we will record position and orientation with respect to the camera frame.
The ROS::tf2 standard message will be used in order to exchange information
between the acquisition process and the other algorithms we will implement. This
choice is due to ROS3 as common framework in this work. An example of the
3D model obtained from the tracker is showed in Figure 2. The OpenNI skeleton
tracker is not open source and it will be probably replaced with a different one,
in order to allow us to modify some internal parameters and improve the system.
1
http://openni.org/
2
http://www.ros.org/wiki/tf/
3
http://www.ros.org/
Fig. 2. 3D skeleton model (joints data obtained using OpenNI tracker). Each coordi-
nates system (the RBG axis) represents position and rotation of a skeleton joint with
respect to (yellow arrows) the camera frame.
4 Action classification
The action classification will lean on a probability-based model. Having a proba-
bilistic model allows us to classify actions in a compact and efficient representa-
tion extracting the principal action constraints. This peculiarity is fundamental
in working with actions involving a large number of degree of freedom, like in
the human body. In fact, the demonstrations can vary one from each other and
some movements could be not essential to reach the goal. A probabilistic ap-
proach makes also possible to perform incremental construction of a model, thus
making easier to refine the representation with new demonstrations. The pur-
pose of this work is make the robot able to reproduce the skills learned using
the action recognition system. A regression process will be developed to archive
an effective robot control with smooth trajectories in a fast and analytic way.
The idea is to extract a “mean” trajectory that would change in the “variance”
field of the acquired trajectories depending on the situation. Figure 3 shows the
idea for an example trajectory in 2D space.
(a) (b)
Fig. 3. Example of 2D trajectories extraction (a) and their regression (b).
5 Conclusion
In this paper we presented an human action recognition system designed to
reproduce the acquired skills on a humanoid robot. We plan to test our system
first in a simulated environment, and then with two different humanoids: VStone
Robovie-X and Aldebaran Nao (Figure 4). Robovie-X is an affordable small robot
perfect to test robustness and flexibility. Nao is bigger than Robovie-X and it
can mount on it the RGB-D sensor to perform real-time tests on a moving robot.
(a) (b)
Fig. 4. Robot proposed for our tests: Robovie-X (a) and Nao (b).
References
1. Bobick, A.F., Davis, J.W.: The recognition of human movement using temporal
templates. Pattern Analysis and Machine Intelligence, IEEE Transactions on 23(3),
257–267 (Aug 2002)
2. Brand, M.: Understanding manipulation in video. In: Proceedings of the 2nd Inter-
national Conference on Automatic Face and Gesture Recognition (FG ’96). pp. 94–.
FG ’96, IEEE Computer Society, Washington, DC, USA (1996)
3. Calinon, S.: Robot Programming by Demonstration: A Probabilistic Approach.
EPFL/CRC Press (2009)
4. Gong, S., Xiang, T.: Recognition of group activities using dynamic probabilistic
networks. In: Proceedings of the Ninth IEEE International Conference on Computer
Vision - Volume 2. pp. 742–. ICCV ’03, IEEE Computer Society, Washington, DC,
USA (2003)
5. Meltzoff, A.N., Moore, M.K.: Imitation of Facial and Manual Gestures by Human
Neonates. Science 198(4312), 75–78 (Oct 1977)
6. Moënne-Loccoz, N., Brémond, F., Thonnat, M.: Recurrent bayesian network for
the recognition of human behaviors from video. In: Proceedings of the 3rd interna-
tional conference on Computer vision systems. pp. 68–77. ICVS’03, Springer-Verlag,
Berlin, Heidelberg (2003)
7. Schaal, S.: Learning from demonstration. In: Advances in Neural Information Pro-
cessing Systems 9 (1997)
8. Siciliano, B., Khatib, O. (eds.): Springer Handbook of Robotics. Springer (2008)
9. Turing, A.M.: Computing machinery and intelligence (1950)
10. Wei, W., Yunxiao, A.: Vision-based human motion recognition: A survey. In: Pro-
ceedings of the 2009 Second International Conference on Intelligent Networks and
Intelligent Systems. pp. 386–389. ICINIS ’09, IEEE Computer Society, Washington,
DC, USA (2009)
11. Wolpert, D.M., Ghahramani, Z., Jordan, M.I.: An internal model for sensorimotor
integration. Science 269, 1880–1882 (1995)