Contextual Clustering for Automated State Estimation by Sensor Networks

Through the lens of an application in space object tracking, we develop the concept of a data-aware algorithmic similarity kernel that enables clustering of partial state observations according to their source phenomena. Particularly, we consider data sets consisting of such observations made by distributed sensor networks where pairwise comparisons yield no useful association. Utilizing the data set as context, likelihoods of association for pairs are assigned by an algorithmic similarity measure that incorporates expert knowledge and domain specific heuristics through statistical analysis of higher order tuples. Spectral clustering is applied to the resulting affinity matrix to partition the data by source, where further expert analysis can better characterize the states being observed. The example application has low dimensionality and relatively simple statistical associations that make it an ideal model for illustrating the overall approach.