Scalable topological and geometric methods for multimodal activity modeling
Abstract
Understanding and modeling dynamical phenomenon is often critical in various applications; from surveillance, intervention, to forensic analysis. Dynamical events in general are difficult to model, exacerbated due to their evolution at multiple spatial-scales, temporal-scales, and non-uniqueness of model parameters that can explain the same observations. Further, in many cases of interest, one encounters observations in high-dimensions, and lack of sufficient observed data to allow effective inferences to be drawn. In this proposal, we are interested in dynamical modeling as applied to human activity analysis from diverse data sources, including visual imagers, wearable devices, and depth sensors. There is a wide disparity in techniques for dealing with the sensor data from different sources, however, we leverage a deeper commonality in mathematical representation. We propose that many of these representations from diverse sensors and data sources can be studied in a unified manner using techniques from geometry and topology. We argue that topological principles in conjunction with geometric metrics can be used to provide a multiplicative effect on the success of state-of-the-art machine learning techniques. The opportunities for applying topological and geometric methods in data and model understanding/inference are tremendous, but this area is under-researched. In this context, we propose the following three thrusts. Task A: Leveraging topology and geometry for dynamical modeling. In this task, we will investigate two related approaches: one focuses on modeling topological properties of dynamical attractors using geometric-metrics, the other focuses on generalizing classical dynamical invariants to cases where the domain of evolution has a known manifold- structure. Task B: Task-adaptive representation reduction. In this task, we consider representations and algorithms that allow one to develop effective algorithms for low bit-rate coding and matching, for sequences evolving on manifolds. These methods can form the basis for resource-constrained deployment for a variety of activity analysis applications. Task C: Applications to human activity modeling and prediction. In this task we apply the techniques for dynamical modeling developed in earlier thrusts, in conjunction with low-dimensional coding techniques for the purposes of activity recognition, summarization, and prediction.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 16, 2018
- Source ID
- W911NF1710293
Entities
People
- Pavan Turaga
Organizations
- Arizona State University
- Army Contracting Command
- United States Army