Semester

Fall

Date of Graduation

2013

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Gianfranco Doretto

Committee Co-Chair

Donald Adjeroh

Committee Member

Mark Culp.

Abstract

Recognizing human activities from video is an important step forward towards the long-term goal of performing scene understanding fully automatically. Applications in this domain include, but are not limited to, the automated analysis of video surveillance footage for public and private monitoring, remote patient and elderly home monitoring, video archiving, search and retrieval, human-computer interaction, and robotics. While recent years have seen a concentration of works focusing on modeling and recognizing either group activities, or actions performed by people in isolation, modeling and recognizing binary human-human interactions is a fundamental building block that only recently has started to catalyze the attention of researchers.;This thesis introduces a new modeling framework for binary human-human interactions. The main idea is to describe interactions with spatio-temporal trajectories. Interaction trajectories can then be modeled as the output of dynamical systems, and recognizing interactions entails designing a suitable way for comparing them. This poses several challenges, starting from the type of information that should be captured by the trajectories, which defines the geometry structure of the output space of the systems. In addition, decision functions performing the recognition should account for the fact that the people interacting do not have a predefined ordering. This work addresses those challenges by carefully designing a kernel-based approach that combines non-linear dynamical system modeling with kernel PCA. Experimental results computed on three recently published datasets, clearly show the promise of this approach, where the classification accuracy, and the retrieval precision are comparable or better than the state-of-the-art.

Share

COinS