Phase II year
2021
(last award dollars: 1685457600)
Phase II Amount
$1,000,000
Intelligence analysts attempting to geolocate ground-level imagery rely on numerous sources of information including reference data of distinct urban and topographical structures which of may be visible from the ground. This matching process is tedious and requires careful identification of topographic features (e.g. ridgelines) and relevant contextual information (e.g. environment type) from both a ground-view and aerial perspective. An automated technique to perform feature extraction, matching and geolocalization will dramatically reduce the workload of the analyst and provide useful analytical information for other intelligence tasks. \n\n Recent research in cross-view image retrieval has shown that distinctive features in ground-level imagery can be correlated with reference datasets to localize the position and orientation of a ground-view sensor. By applying techniques in multi-modal cross-view learning it is possible to correlate multiple visual concepts between aerial and ground-level imagery. The resulting embeddings can then be used for efficient search using dimensionality reduction techniques for generating search indices. To perform geolocalization, these indices can then be queried in real time using a modular and parallelized processing framework and database.
