Efficient Condensation of Spatial/Temporal Data

9971784

The proposed research concerns the condensation and analysis of spatial/temporal data for purposes of classification and anomaly detection. The focus is a new approach to landmark estimation and spatial clustering based on spatial point processes. Such processes provide a potentially rich class of models to express high-level prior knowledge about curves and shapes. These models are more suitable than commonly used discrete Markov random field models and template deformation models, especially in situations where there is a lack of training data. A version of Bayesian nonparametric curve estimation will be developed using the new approach. A principal objective is to achieve a parsimonious specification of the landmark information (describing the location of the knots in a spline curve for example). A further objective is improved methods for simultaneous inference from temporal data.

The benefit of an efficient statistical model is data condensation, or the reduction of often unmanageably large data sets to a parsimonious form, without the sacrifice of key statistical information. Another benefit is that it can lead to the discovery of interesting anomalies, given the availability of suitable inferential methods to lend credence to such findings. Specific applications to be explored include off-line signature recognition and 2D-gel electrophoresis imaging.