Statistical Models and Methods for Video Analytics in Camera Networks

Visual sensor networks, also known as camera webs, are becoming increasingly prevalent with 30 million surveillance units deployed in the US today producing 4 billion hours of footage per week (Popular Mechanics, Jan. 2007). Since visible-light cameras provide excellent spatio-temporal resolution, long range, wide field of view and low latency, they hold a great potential for pervasive wide-area monitoring. However, the processing of this rich source of data will become increasingly unsustainable unless autonomous (no human operator in the loop) camera systems and networks are developed capable of handling  highly cluttered indoor as well as outdoor environments.

The goal of this research thrust is to develop statistical models and methods for visual data analysis in a networked setting. Of particular interest is the development of techniques for abnormal behavior detection, such as motor accident, stalled vehicle, abandoned suitcase, etc. Below are listed some ongoing projects.

Probabilistic methods for adaptive background subtraction (2007-...)
Team: J.M. McHugh, J. Konrad, P.-M. Jodoin, V. Saligrama, D. Castanon
Funding: Air Force Office of Sponsored Research (SBIR), College of Engineering Catalyst Award
Background subtraction by means of probability thresholding has been mature for a number of years. We have revisited this approach, however, from the standpoint of binary hypothesis testing, and have added two improvements. First, we have enriched the background model with a foreground model estimated from spatial samples under spatial ergodicity assumption. Furthermore, we have incorporated a label model based on Markov random fields in order to assure spatial continuity of foreground labels detected. The proposed methodology is independent of specific probabilistic models for the background and foreground; we have used a non-parametric kernel model (Parzen window) with Gaussian kernel.

Behavior subtraction (2007-...)
Team: P.-M. Jodoin, V. Saligrama, J. Konrad
Funding: College of Engineering Ignition Award
Although background subtraction may be sufficient for restricted-area access control, it is not sufficient for the detection of unusual behavior. Thus, the question one needs to answer is: How to distinguish normal motion patterns from abnormal ones?  Using the background detection results, such as those shown above, we have developed an abnormal behavior detection method based on the maximum activity principle. By aggregating motion labels for a given pixel from all frames of a training sequence, we build a low-dimensionality representation against which we test an aggregate computed from the observed sequence. If the observed aggregate exceeds the training aggregate, abnormality is declared. This simple strategy leads to remarkably robust results.

.