Project Title: Mining, Indexing, and Retrieval of Spatio-Temporal Patterns in Video Databases of Human Motion

National Science Foundation Award Number: IIS-0308213 (09/15/03 -- 09/14/06)

Stan Sclaroff, PI
George Kollios and Margrit Betke, Co-PI's
Department of Computer Science
Boston University
111 Cummington St., Boston, MA 02215, U.S.A.
Phone: (617) 353-8919, Fax: (617) 353-6457
E-mail: {sclaroff,gkollios,betke}@cs.bu.edu
URL: http://www.cs.bu.edu/fac/sclaroff
URL: http://www.cs.bu.edu/fac/gkollios
URL: http://www.cs.bu.edu/fac/betke

Keywords:

spatio-temporal data mining, motion time series data, video database methods

Project Summary:

The aim of this research project is to develop and test methods for indexing, retrieval, and data mining of human motion trajectories in video databases. Computer vision techniques are being devised for automatic extraction of human motion time series data from video. Data mining algorithms are being developed that can be used to discover clusters and other patterns in the extracted motion time-series data. One promising direction being explored is to model the observed motion time series sequences with a finite mixture of Hidden Markov Models (HMMs). Use of the HMM representation presents certain advantages with regard to modeling; however, it presents important challenges for the design of efficient clustering, indexing, and retrieval algorithms. Thus more efficient, sampling-based and embedding-based methods must be formulated. The resulting ideas are evaluated in a prototype video retrieval system, with real-world video datasets that depict human body motion. Synthetic sequences (e.g., generated via computer graphics) are used in quantitative performance experiments where ground truth information is required.

Publications and Products:

There is nothing to report yet, since this is a new award.

Project Impact:

Data mining and indexing methods for video databases of human motion would enable numerous applications that are valuable to society, for instance:
  1. Video indexing, mining, and retrieval of archival footage. A large percentage of video transmission depicts human activities: sports, news, entertainment, education, etc.
  2. Assistive environments. Video data mining of human motion could be useful in homecare for the handicapped, infirm, elderly; e.g., a person's day to day motion patterns could be analyzed for illness or decline and alert caregivers. Such an approach could also prove useful in occupational safety.
  3. Homeland Security. Particularly useful would be the ability to cluster and retrieve common motion patterns, and to employ the models gained in data mining for outlier detection thereby alerting database analysts to anomolous activities discovered in the video data streams.
  4. Video-based analysis of human motion. Data mining methods could enable automated studies of human movement. Such motion analysis should prove to be a very useful tool in occupational safety and therapy, as well as in training for sports and dance.

Goals, Objectives, and Targeted Activities:

This effort will focus on indexing and data mining methods for video databases of human motion. Such data has a spatio-temporal askect that must be dealt with, and therefore a major focus of the project will be on developing efficient methods for indexing, clustering and mining databases of motion time-series data.   Another important part of the effort will focus on methods that can automatically extract and analyze motion time-series from video databases.  

Area Background:

Up until very recently, data mining of video databases of human motion received only a small amount of attention [1], while there has been a great deal of related work in video-based surveillance and activity monitoring (too numerous to mention here) .  Recently there have been two workshops organized to explore potential directions for research on this topic, for instance see [2,3].  On the other hand, a variety of methods for clustering time-series data have been proposed recently, particularly in the statistics and data mining communities.  For example, the Fourier transform has been successfully applied in clustering certain types of time-series data [4], splines have been used to cluster functional data [5].  In other work, Gaffney, et al. [6] proposed trajectory clustering with a mixture of regression models, where the regression models can be either parametric (e.g., linear, quadratic) or non-parametric (e.g., Gaussian Kernels).  Cadez, et al, [7] gave a general EM-based probabilistic framework for clustering non-vector data.   A number of HMM-based clustering methods have also been proposed ([8,9.10] to name a few).  

Area References

[1] G. Kollios, S. Sclaroff, and M. Betke.  Motion mining: Discovering spatio-temporal patterns in databases of human motion. In Proc. ACM-SIGMOD Int. Workshop on Data Mining and Knowledge Discovery (DMKD), May 2001.

[2] A. Rosenfeld, D. DeMenthon, and D. Doermann, eds. Proc. DIMACS Workshop on Video Mining, 2002.

[3] I. Haritaoglu, and T. Syeda-Mahmood, eds., Proc. IEEE Workshop on Event Mining: Detection and Recognition of Events in Video, 2003.

[4] R. Agrawal, C. Faloutsos, and A. Swami.  Efficient similarity search in sequence databases, in Proc. Int. Conf. on Foundations fo Data Organization and Algorithms, pp. 69-84, 1993.

[5] G. James and C. Sugar, Clustering for sparsely sampled functional data, Technical Report, Dept. of Information and Operations Management, USC, 2002.

[6] S. Gaffney and P. Smyth. Trajectory clustering with mixtures of regression models. In Knowledge Discovery and Data Mining, pp. 63-72, ACM Press, 1999.

[7] I.V. Cadez, S. Gaffney, and P. Smyth.  A general probablistic framework for clustering individuals and objects. In Knowledge Discovery and Data Mining, pp. 140-149, ACM Press, 2000.

[8] P. Smyth. Clustering sequences with hidden Markov models.  In Advances in Neural Information Processing Systems (NIPS), vol. 9, pp. 648-654, MIT Press, 1997.

[9] C. Li and G. Biswas. A Bayesian approach to temporal data clustering using hidden Markov models. In Proc. International Conf. on Machine Learning, pp. 599-608, 2000.

[10] T. Oates, L. Firoiu, and P.R. Cohen. Using dynamic time warping to bootstrap HMM-based clustering of time series.  In Sequence Learning: Paradigms, Algorithms, and Applications. Springer, 2000.

Potential Related Projects:

These will be posted here later.

Project Web site URL:

http://www.cs.bu.edu/groups/ivc/MotionMining/

Online software:

There is nothing to report yet, since this is a new award.

Online resources:

There is nothing to report yet, since this is a new award.