REPOSITORY > RESULTS

Abstract

Many applications, including smart environments, surveillance, human-robot interaction,
and ambient assisted living, involve the problem of learning patterns of agent behavior from
sensor data. Deviant behavior is a pattern in the data that either does not conform to the
expected behavior, that is, anomalous behavior, or matches previously defined unwanted
behavior, that is, suspicious behavior. The present thesis focuses on the detection of behavior
patterns representing a security risk, health problem, or other abnormal behavior
contingency.
Real-life applications for deviant behavior detection present several challenges. First,
plan recognition research has assumed that atomic actions are either given or can be trivially
obtained, while real-life applications require activity recognition from raw sensor readings.
The second challenge is how to
exibly encode complex, unstructured, daily-living behavior
patterns that do not follow predefined scenarios. Thirdly, deviant behavior may be reflected
on different time scales and different modalities, which raises the question of how to combine
different time scales and modalities into a single evaluation. Finally, many domains include
behavior in which no single event is sufficient to decide whether the behavior is deviant;
therefore, an advanced approach is required to accumulate deviation over time.
This thesis proposes a unified framework to analyze agent behavior from prior knowledge
and external observations in order to detect deviant behavior patterns, regardless of whether
the observed entities are humans, software agents, or even robots. To address the problem of
activity recognition from sensor data, the thesis introduces an activity recognition pipeline
that includes filtering, attribute construction, activity identification, and activity smoothing.
From the behavior analysis perspective, we propose a novel, efficient encoding that we
refer to as a spatio-activity matrix. This matrix is able to capture behavior dynamics in a
specific time period using spatio-temporal features, whereas its visualization allows visual
comparison of different behavior patterns. The thesis also provides a feature extraction
technique, based on principal component analysis, in order to reduce the dimensionality of
the spatio-activity matrix. We then introduce a clear problem definition that helps establish
a theoretical framework for detecting anomalous and suspicious behavior from agent traces
in order to show how to optimally perform detection. We discuss why detection error is
often inevitable and prove the lower error bound, and provide several heuristic approaches
that either estimate the distributions required to perform detection or to directly rank the
behavior signatures using machine learning approaches. The established theoretical framework
is extended to show how to perform detection when the agent is observed over longer
periods of time and no significant event is sufficient to reach a decision. We specify conditions
that any reasonable detector should satisfy, analyze several detectors, and propose a
novel approach, referred to as a F-UPR detector, that generalizes utility-based plan recognition
with arbitrary utility functions. The unified framework is demonstrated empirically
in three studies. The first study addresses detection of decreased behavior that indicates
disease or deterioration in the health of elderly persons, while the second study deals with
the detection of suspicious passengers in the airport simulation. Finally, the third study
concerns the verification of persons at an access control point in high-security application.