Models and data sources for economical computer vision in camera networks

Snyder, Corey Ethan

Permalink

https://hdl.handle.net/2142/125616 Copy

Description

Title

Models and data sources for economical computer vision in camera networks

Author(s)

Snyder, Corey Ethan

Issue Date

2024-07-11

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

Do, Minh N

Doctoral Committee Chair(s)

Do, Minh N

Committee Member(s)

• Boppart, Stephen
• Liang, Zhi-Pei
• Schwing, Alexander
• Shomorony, Ilan

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Computer vision
• machine learning
• signal processing
• camera networks
• pattern recognition
• cost-constrained machine learning

Abstract

The past decade of machine learning and computer vision research has led to landmark achievements in numerous tasks from image classification to image generation to panoptic segmentation. These advancements are predicated on the ``Big Data'' paradigm where exceptionally large deep learning models, annotated datasets, and expensive computational infrastructure enables such solutions. In this dissertation, we look to settings where model size, amount of data, and computational cost are constrained. We refer to such settings as ``economical computer vision''. In particular, we focus on camera networks as our target environment and motivate the need for economical solutions for this setting. We present two datasets, STREETS and its extension STREETS BFS, as highly practical demonstrations of how camera networks may be used for intelligent transportation systems across a large suburban county. We shape our development of both datasets through conversations with traffic engineers who operate this network and present benchmark tasks and experiments directed at economical solutions for these engineers. We also propose computer vision models aimed at efficient and effective solutions for the task of background foreground separation (BFS). We introduce our RUSTIC model, the first known algorithm to combine radar and camera modalities, as a multi-model and unsupervised solution for BFS. RUSTIC employs the technique of algorithm unrolling to yield a fast fusion of radar data within the Robust PCA framework. The impressive multi-scene generalization of RUSTIC then motivates our proposed GrayNet model which combines white-box algorithm unrolling with popular black-box deep learning models. We see through rigorous empirical evaluation and ablation studies that GrayNet is a highly label-efficient, interpretable, and fast solution for BFS that may enable economical computer vision in camera networks as well as motivate a variety of future works that extend the work of this dissertation.

Graduation Semester

2024-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/125616

Copyright and License Information

Copyright 2024 Corey Snyder

Owning Collections