Small Sample Size Semi-Supervised Feature Clustering for Detection and Classification of Objects and Activities in Still and Motion Multi-spectral Imagery

Toyon Research Corp. and the Penn State Univ. propose research and development of innovative algorithms for classifying objects and activities observed in high-dimensional data extracted from multi-sensor motion imagery. The proposed algorithms include novel feature clustering techniques to enable effective characterization of intra-class and inter-class appearance variations in datasets containing a small number of labeled, and a large number of unlabeled, high-dimensional feature vectors. The proposed development is expected to provide significant improvements in object and activity classification performance, including maximization of the probability of correct classification and minimization of false declaration rates for real-world applications including highly variable clutter and object and activity types not represented amongst the labeled training data. The proposed algorithmic framework is of a general nature and utility, and will be demonstrated for multiple practical applications in Phase II. Integration in AFRL systems will be supported in Phase II, and related transition opportunities will be pursued.