Improved robustness and efficiency for automatic visual site monitoring

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2009.This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.Cataloged from student-submitted PDF version of thesis.Includes bibliographical references (p. 219-228).Knowing who people are, where they are, what they are doing, and how they interact with other people and things is valuable from commercial, security, and space utilization perspectives. Video sensors backed by computer vision algorithms are a natural way to gather this data. Unfortunately, key technical issues persist in extracting features and models that are simultaneously efficient to compute and robust to issues such as adverse lighting conditions, distracting background motions, appearance changes over time, and occlusions. In this thesis, we present a set of techniques and model enhancements to better handle these problems, focusing on contributions in four areas. First, we improve background subtraction so it can better handle temporally irregular dynamic textures. This allows us to achieve a 5.5% drop in false positive rate on the Wallflower waving trees video. Secondly, we adapt the Dalal and Triggs Histogram of Oriented Gradients pedestrian detector to work on large-scale scenes with dense crowds and harsh lighting conditions: challenges which prevent us from easily using a background subtraction solution. These scenes contain hundreds of simultaneously visible people. To make using the algorithm computationally feasible, we have produced a novel implementation that runs on commodity graphics hardware and is up to 76 faster than our CPU-only implementation. We demonstrate the utility of this detector by modeling scene-level activities with a Hierarchical Dirichlet Process.(cont.) Third, we show how one can improve the quality of pedestrian silhouettes for recognizing individual people. We combine general appearance information from a large population of pedestrians with semi-periodic shape information from individual silhouette sequences. Finally, we show how one can combine a variety of detection and tracking techniques to robustly handle a variety of event detection scenarios such as theft and left-luggage detection. We present the only complete set of results on a standardized collection of very challenging videos.by Gerald Edwin Dalley.Ph.D