Sensing and Efficient Inference for Identity Management
Abstract
We developed a mathematical formulation and a set of algorithms that make significant strides towards a practical system for identity management. At the core of the formulation lies a single binary integer program that describes the key data association problem: Nodes in a graph correspond to observations, and edges are weighted with correlation measures that quantify positive or negative evidence for the hypothesis that two nodes correspond to observations of the same person. The binary integer program defines a partition of the nodes into sets that are meant to correspond to distinct identities. Solving this problem is NP-hard, and we developed a problem decomposition method that, while losing optimality guarantees, show good empirical performance at near frame-rate. To evaluate our method and establish a baseline for future work by us and others, we developed a large video data set with more than 1 million frames and more than 2000 identities observed from eight cameras placed on the campus of Duke University. The data set is fully annotated, and a 3D trajectory is available for each person in every frame from every camera. We also formulated a new methodology for performance evaluation in identity management.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 20, 2015
- Accession Number
- ADA631617
Entities
People
- Carlo Tomasi
Organizations
- Duke University