Tools for acquired scene analysis and transmission

Kamali Moghaddam, Mahsa

Permalink

https://hdl.handle.net/2142/29493 Copy

Description

Title

Tools for acquired scene analysis and transmission

Author(s)

Kamali Moghaddam, Mahsa

Issue Date

2012-02-01T00:49:09Z

Director of Research (if dissertation) or Advisor (if thesis)

Hart, John C.

Committee Member(s)

• Nahrstedt, Klara
• Campbell, Roy H.
• Ofek, Eyal
• Stroila, Matei

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Date of Ingest

2012-02-01T00:49:09Z

Keyword(s)

• Scene Analysis
• Scene Acquisition
• 3D Data Analysis
• Image Morphing
• LiDAR Data
• 3D Point Cloud
• Tele Immersion
• 3D Data Classification

Abstract

Recent advancement in research and technology on virtual environments has significantly escalated users' expectations on this matter. This trend is easily witnessed, for example, by websites like twitter or facebook. Moreover, searching for 3D navigation in any search engine can provide an increasing number of instances where the full 3D models of cities are available to users for a virtual walk through. From research viewpoint, design of such virtual environments needs to address two key concepts of “user interface” and “data manipulation”. A wide spectrum of user interfaces has entered into the industry of virtual navigation. These range from haptic devices such as joysticks, Wii, and touch screens to the more sophisticated options such as Microsoft Kinect. Obviously, the potential of these devices needs to be carefully evaluated in each application to strike a balance between efficacy and complexity. In this thesis, we have investigated a few case studies on quite different tasks to justify the universality of this issue. The tasks vary from simulated circuit design to GPS navigation. Of particular interest is the task of choreography software, where the interface evolved from dancers' personal experience into an optimal combination of preset virtual scenes and real time control of the scene by a Wii device. Data manipulation methods vary depending on whether 2D or 3D data is being used. The choice between 2D versus 3D depends on the available bandwidth and processing power. For example, when working with navigation maps on mobile devices, one cannot easily transmit huge 3D data. Hence, researchers tend to either use graphical models or 2D images. On the other hand, when high bandwidth and processing power is available, it is often better to use full 3D capabilities. Regardless of the choice of 2D or 3D sensor, there are certain limitations for each one. For example, 2D sensors often suffer from limited field of view which is a major strain when user wants to get a sense of the entire scene. On the other hand, while 3D sensors can potentially provide 360 degrees view of the scene, their captured data is too noisy and sparse to be useful for any automatic structure detection. In this thesis, we address both 2D and 3D data and develop novel techniques to remedy the aforementioned problems within specific applications. The ultimate results of these works provide robust enhancement and detection in 2D and 3D data. We briefly overview the 2D and 3D tasks here and leave the details to their associated chapter in this document. We study two 2D tasks focusing on issues of robust detection and panoramas creation respectively. The first task, which we named Meth Morph, analyzes 2D portrait images of human face and morphs them into a face as if the subject is drug (Meth) addicted. The developed technique requires only a single input image and performs the rest of the task in a fully automatic fashion. The robustness of the systems is perhaps best justified by referring to the successful live demo given to random visitors. This thesis also investigates overcoming the limited field of view in 2D cameras by available panorama creation methods. A very beneficial application for panoramas is being able to view a long street in one single view. For example, when working with mobile devices, one needs a highly efficient way to transmit and display a full length street. This was made possible by a project called Street Slide at Microsoft. Nevertheless, clutter over the pace of stitched images (e.g. telephone wires) can seriously damage the performance of street panoramas. To improve the robustness of panorama creating, in this thesis, we develop a fully automatic detection and removal of wire-like clutters in the images. Finally, we investigate robustness issues in 3D data processing and structure detection. The first challenge with processing 3D data is their huge size compared to 2D. This makes their processing computationally expensive; an issue that is in conflict with their real-time use as in interactive environments. For this purpose, we develop an efficient system based on an OCtree spatial data structure allowing real-time rendering and manipulation of a 3D point cloud. The other challenge in processing 3D data is the issues of noise and outliers. In this thesis, we develop a novel method for robust classification of curvilinear and surface like structures within a point cloud. All the mentioned pieces help improve the efficiency and quality of data processing, analysis and even transmission. The latter is of special demand by nowadays technology of mobile devices.

Graduation Semester

2011-12

Permalink

http://hdl.handle.net/2142/29493

Copyright and License Information

© 2011 by Mahsa Kamali Moghaddam. All rights reserved.

Owning Collections