13-06-2012, 03:45 PM
Recognizing and Remembering Individuals:
Online and Unsupervised Face Recognition for Humanoid Robot
[attachment=24363]
Abstract
Individual recognition is a widely reported phenomenon
in the animal world, where it contributes to successful
maternal interaction, parental care, group breeding,
cooperation, mate choice, etc. This work addresses the
question of how one may implement such social
competence in a humanoid robot.
Introduction
The ability to recognize and remember individuals is
crucial for complex interactions among social animals,
such as preferential treatment, cooperative behavior, and
reciprocity. Mantis shrimps have been observed to avoid
empty cavities with the odor of those that have defeated
them in the past, but enter those with the odor of
individuals they have beaten. Caldwell [2] concluded
that they’re able to not only recognize other individuals,
but also remember their reputation as fighters.
Related Work
Research in person identification technology has recently
received significant attention, due to the wide range of
biometric, information security, law enforcement
applications, and Human Computer Interaction (HCI).
Face recognition is the most frequently explored
modality and has been implemented using various
approaches [5]. A combination of face and body
recognition has been proposed [9]. Speech recognition
has also been widely investigated [17]. The use of
multiple modalities has been observed [7,8].
Design Issues and Considerations
Performance Criteria
Current state of the art in face recognition technology
allows for a recognition accuracy of 95% on more than
1000 frontal mugshot-like images when taken in the
same day and 80% when taken with a different camera
and lighting condition [11]. Our performance criteria,
however, is less ambitious in terms of recognition
accuracy per image. Our goal is for the system to be able
to consistently recognize and remember people who are
relevant to Kismet and interact with it on a regular basis.
In the same way, we also do not remember every single
person we pass on the street.
Failure Modes
We consider two possible failure types: clustering error
and failure to learn to recognize a person despite
frequent encounters. While some errors are unavoidable,
it is less harmful to place an individual’s faces into
multiple clusters than it is to cluster multiple individuals
into one class in the training set. In the latter case, the
robot will constantly treat multiple people as the same
person and any effort to learn additional characteristics
of these people will be misleading.
Eigenface Method
We decided to implement the face recognition module
using the eigenface technique [5] because it is widely
implemented and well known for its simplicity and
computational efficiency. We plan to explore other types
of representations as well. The eigenface method uses an
information theory approach of coding facial images,
where it attempts to find the principal component of the
distribution of faces, or the eigenvectors of the
covariance matrix of the set of face images.
Online and Unsupervised Face Recognition for Humanoid Robot
[attachment=24363]
Abstract
Individual recognition is a widely reported phenomenon
in the animal world, where it contributes to successful
maternal interaction, parental care, group breeding,
cooperation, mate choice, etc. This work addresses the
question of how one may implement such social
competence in a humanoid robot.
Introduction
The ability to recognize and remember individuals is
crucial for complex interactions among social animals,
such as preferential treatment, cooperative behavior, and
reciprocity. Mantis shrimps have been observed to avoid
empty cavities with the odor of those that have defeated
them in the past, but enter those with the odor of
individuals they have beaten. Caldwell [2] concluded
that they’re able to not only recognize other individuals,
but also remember their reputation as fighters.
Related Work
Research in person identification technology has recently
received significant attention, due to the wide range of
biometric, information security, law enforcement
applications, and Human Computer Interaction (HCI).
Face recognition is the most frequently explored
modality and has been implemented using various
approaches [5]. A combination of face and body
recognition has been proposed [9]. Speech recognition
has also been widely investigated [17]. The use of
multiple modalities has been observed [7,8].
Design Issues and Considerations
Performance Criteria
Current state of the art in face recognition technology
allows for a recognition accuracy of 95% on more than
1000 frontal mugshot-like images when taken in the
same day and 80% when taken with a different camera
and lighting condition [11]. Our performance criteria,
however, is less ambitious in terms of recognition
accuracy per image. Our goal is for the system to be able
to consistently recognize and remember people who are
relevant to Kismet and interact with it on a regular basis.
In the same way, we also do not remember every single
person we pass on the street.
Failure Modes
We consider two possible failure types: clustering error
and failure to learn to recognize a person despite
frequent encounters. While some errors are unavoidable,
it is less harmful to place an individual’s faces into
multiple clusters than it is to cluster multiple individuals
into one class in the training set. In the latter case, the
robot will constantly treat multiple people as the same
person and any effort to learn additional characteristics
of these people will be misleading.
Eigenface Method
We decided to implement the face recognition module
using the eigenface technique [5] because it is widely
implemented and well known for its simplicity and
computational efficiency. We plan to explore other types
of representations as well. The eigenface method uses an
information theory approach of coding facial images,
where it attempts to find the principal component of the
distribution of faces, or the eigenvectors of the
covariance matrix of the set of face images.