Structured and Collaborative Geometric Signal Models for Big Data Analysis: Theory and Applications in Image, Video and Audio RA 3.1.1
Abstract
Efficient data modeling is critical for ill-posed problems in signal reconstruction, classification, and identification. This project develops a fundamental framework for structured and collaborative design of such signal models. The underlying foundations are derived via a combination of theories from sparse modeling, Gaussian Mixture Models, and Principal Component Analysis. Efficient computational approaches are an intrinsic part of the project as well. In addition to theoretical and computational questions, particular applications here addressed include anomaly detection in videos with dynamic background, collaborative object detection and classification, source identification and separation, activity clustering in video, and dimensionality reduction onto physical meaningful spaces. The combination of multimodalities and incorporation of prior knowledge and side information will be considered as well. Scenarios with significantly under sampled or missing data are addressed as well as part of the proposed frameworks. The theoretical foundations include the development of a novel approach of collaborative compressed sensing and the incorporation of information theory tools into the sparse modeling world. Finally, in order to address critical big data challenges, on-line and real time optimization techniques will be developed. The project is carried out in close collaboration with Department of Defense researches, including frequent visits and technology transfers.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 24, 2017
- Accession Number
- AD1108942
Entities
People
- Guillermo Sapiro
Organizations
- Duke University