Efficient Probabilistic Models for Spatiotemporal Inference

dc.contributor.advisorGogate, Vibhav
dc.contributor.advisorFei, Baowei
dc.contributor.committeeMemberChandrasekaran, R.
dc.contributor.committeeMemberRuozzi, Nicholas
dc.contributor.committeeMemberNatarajan, Sriraam
dc.creatorRoy, Chiradeep
dc.date.accessioned2022-11-29T23:37:12Z
dc.date.available2022-11-29T23:37:12Z
dc.date.created2022-05
dc.date.issued2022-05-01T05:00:00.000Z
dc.date.submittedMay 2022
dc.date.updated2022-11-29T23:37:13Z
dc.description.abstractSequential modeling of spatiotemporal data has a wide range of use cases, such as traffic control modeling, pedestrian trajectory prediction, and activity recognition in videos. Probabilistic models are a good choice for modeling such data because they provide a rich representation for modeling complex, multimodal probability distributions. In addition, they are capable of answering a broad range of interesting probabilistic queries. Despite their many advantages, however, traditional Probabilistic Graphical Models (PGMs) such as Bayesian Networks (BNs) and Markov Networks (MNs) pose problems at inference time because exact inference over them is NP-hard in general. The dynamic and multimodal nature of spatiotemporal sequence data further compounds this problem. This dissertation explores models and methodologies that admit either fully tractable inference or efficient approximate inference queries over sequential spatiotemporal data to address these problems. More specifically, we make the following contributions: • We introduce AND/OR Conditional Cutset Networks (AOCCNs), a discriminative tractable probabilistic model (TPM) that admits polynomial-time inference for computing point-wise conditional probabilities. • We introduce the Dynamic Cutset Network (DCN) framework – an extension of the Cutset Network framework to the temporal domain – that allows for efficient approx- imate inference using particle filtering in the general case and exact inference when certain structural constraints are obeyed. • We show how the DCN framework can be combined with neural networks to create a powerful video activity recognition system to enhance prediction accuracy and increase the transparency of the functioning of the joint model by providing explanations to probabilistic queries. • We introduce a fully generative, dynamic, grid-based, probabilistic framework called the Probabilistic Grid Modeller (PROGRIMO) to efficiently model and reason about trajectory data on a fixed-sized grid.
dc.format.mimetypeapplication/pdf
dc.identifier.uri
dc.identifier.urihttps://hdl.handle.net/10735.1/9550
dc.language.isoen
dc.subjectComputer Science
dc.titleEfficient Probabilistic Models for Spatiotemporal Inference
dc.typeThesis
dc.type.materialtext
thesis.degree.collegeSchool of Engineering and Computer Science
thesis.degree.departmentComputer Science
thesis.degree.grantorThe University of Texas at Dallas
thesis.degree.namePHD

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
ROY-PRIMARY-2022-1.pdf
Size:
15.45 MB
Format:
Adobe Portable Document Format

License bundle

Now showing 1 - 2 of 2
No Thumbnail Available
Name:
PROQUEST_LICENSE.txt
Size:
5.84 KB
Format:
Plain Text
Description:
No Thumbnail Available
Name:
LICENSE.txt
Size:
1.84 KB
Format:
Plain Text
Description: