Perceptual data mining : bootstrapping visual intelligence from tracking behavior

Abstract

One common characteristic of all intelligent life is continuous perceptual input. A decade ago, simply recording and storing a a few minutes of full frame-rate NTSC video required special hardware. Today, an inexpensive personal computer can process video in real-time tracking and recording information about multiple objects for extended periods of time, which fundamentally enables this research. This thesis is about Perceptual Data Mining (PDM), the primary goal of which is to create a real-time, autonomous perception system that can be introduced into a wide variety of environments and, through experience, learn to model the activity in that environment. The PDM framework infers as much as possible about the presence, type, identity, location, appearance, and activity of each active object in an environment from multiple video sources, without explicit supervision. PDM is a bottom-up, data-driven approach that is built on a novel, robust attention mechanism that reliably detects moving objects in a wide variety of environments. A correspondence system tracks objects through time and across multiple sensors producing sets of observations of objects that correspond to the same object in extended environments. Using a co-occurrence modeling technique that exploits the variation exhibited by objects as they move through the environment, the types of objects, the activities that objects perform, and the appearance of specific classes of objects are modeled. Different applications of this technique are demonstrated along with a discussion of the corresponding issues.

(cont.) Given the resulting rich description of the active objects in the environment, it is possible to model temporal patterns. An effective method for modeling periodic cycles of activity is demonstrated in multiple environments. This framework can learn to concisely describe regularities of the activity in an environment as well as determine atypical observations. Though this is accomplished without any supervision, the introduction of a minimal amount of user interaction could be used to produce complex, task-specific perception systems.