Digital Library
for Human Movement
PI: Jezekiel Ben-Arie, ECE & CS Depts. , University of Illinois
at Chicago
Contact Information
|
Jezekiel Ben-Arie |
IDM’02
Report: http://www.ece.uic.edu/~benarie/idm02/
Web Based Demo of Digital Video Library: http://vision.ece.uic.edu/cgi-bin/vdsearch.cgi
|
List of Supported Students and Staff Zhiqian Wang, Xiao Du, Purvin Pandit, Suman Puthana, Shyam Sundar Rajaram, Arvind Mohan, Manoj Seshadrinathan, Dayan Sivalingam. Project Award Information Keywords Project Summary This project proposes research towards the
development of a video database for a digital library of human
actions/activities. Analysis of human motion is important in many areas such as
kinesiology, ergonomic research, anthropology, biomechanics, rehabilitation,
sports medicine and athletic analysis. In addition, research in artistic
areas such as dancing, choreography and figure skating also requires analysis
of human motion. There are many libraries that include video and other motion
data (such as dance notation) that can be utilized as sources of raw data.
However, such raw data does not quantify the actual three-dimensional human
motion, which could be quite complex. Obviously, there is a need for a
general method for representation of human action that can be employed to
uniquely specify, store and retrieve such actions in digital libraries. Thus,
our research is focused on developing innovative methods for analysis,
representation and interpretation of human actions and activities for video
and other databases that describe human motion. Since natural languages and symbolic
temporal descriptions are not suited to accurately and quantitatively specify
articulated human activities or actions, and other descriptions of human
motion such as dance notations are useful only in their specific domain, we
developed a suitable representation for articulated human motion. Such
representations were adapted from the area of virtual reality modeling that
was developed for humanoid animations. We introduced a representation for human motion,
which is based on multidimensional vectors that specify the angular positions
and velocities of the human skeletal joints. Since skeletal segments are
rigid, the angular motion of the joints uniquely specifies human motion.
Using this motion representation, it is planned to develop a user-friendly interface
that can accept various query modes such as spatio-temporal specification,
specification by visual examples or by interactive symbolic specification
using motion icons. In the symbolic specification, it is also planned to
learn from methods used for dance notations and to develop similar approach
for general human motion. Another
goal of our research is to develop robust and more flexible indexing methods
that can recognize, reconstruct and retrieve queried human actions from video
and other motion databases. For this purpose we developed a novel method for
view-based recognition of human action/activity from videos. By observing
just a few frames, we can identify the activity that takes place in a video
sequence. The basic idea of our method is that activities can be positively
identified from a sparsely sampled sequence of few body poses acquired from
videos. In our approach, an activity is represented by a set of pose and
velocity vectors for the major body parts (hands, legs and torso) and stored
in a set of multidimensional hash tables. We
developed a theoretical foundation that shows that robust recognition of a
sequence of body pose vectors can be achieved by a method of indexing and
sequencing (RISq) and it requires only few pose vectors (i.e. sampled body
poses in video frames). We find that the probability of false alarm drops
exponentially with the increased number of sampled body poses. So, matching
only few body poses guarantees high probability for correct recognition. Our
approach is parallel i.e. all the possible model activities are examined at
one indexing operation, since all the model activities are stored in the same
set of hash tables. In addition, our method is robust to partial occlusion
since each body part is indexed separately. We use a sequence based voting
approach to recognize the activity invariant to the activity speed.
Experiments performed with videos having 8 different activities show robust
recognition with our method. The method is also robust in conditions of varying
view angle in the range of + 30 degrees. Project Publications and Products ·
An innovative method for human activity recognition was developed.
This method is based on Sequential Indexing and is described in the project
summary. ·
We have succeeded in segmentation of human faces using color and
frequency characteristics. We have completed a model-based segmentation
scheme that detects man-made objects in cluttered scenes. ·
A robust scheme for head detection was developed with successful
results for various poses of human head. Also, a robust scheme for tracking
human body parts for the purpose of activity recognition was constructed. ·
We have developed a web site: http://vision.ece.uic.edu/cgi-bin/vdsearch.cgi
with a library of videos that are retrieved according to the queries. This
web site is only an initial example of the planned digital library. ·
An Interactive GUI for queries on human actions/activities is in
advanced stage of development.
·
Nandy, D. and Ben-Arie, J., “Recovery of 3D Face Structure
using Recognition,” appeared in IEEE-IAPR International Conference on
Pattern Recognition ICPR), Barcelona, Spain, Vol. 1, pp.1104-1108,
September 2000.
Project Impact · Human Resources: Ph.D. Students: Zhiqian Wang, Xiao Du. M.S. Students: Suman Puthana, Purvin Pandit, Shyam Sundar Rajaram, Arvind Mohan, Manoj Seshadrinathan, Dayan Sivalingam, Debra Hill, Naren Pandian, Sowmya Parthan, Praveen Shangunathan. · Education and curriculum development at all levels. The PI, Prof. Ben-Arie teaches graduate and undergraduate courses on image understanding, image analysis and image processing. His research has influenced and contributed to the contents of these courses. Goals, Objectives, and Targeted Activities
Area Background Content Based retrieval from Video Databases is an
active area of research. The main research goal is to facilitate retrieval of
video objects based on their content. The area spans multiple disciplines of
which the Database and Information Retrieval and the Computer Vision and
Image Processing Disciplines. The Database area is concerned about languages
for specifying queries, efficient retrieval mechanisms. The Computer Vision
and Image processing areas are concerned about processing raw videos, and
extracting quantitative and qualitative information from them. Area References
Potential Related Projects |