The multi-level learning and classification of multi-class parts-based representations of U.S. Marine postures

This paper primarily investigates the possibility of using multi-level learning of sparse parts-based representations of US Marine postures in an outside and often crowded environment for training exercises. To do so, the paper discusses two approaches to learning parts-based representations for each posture needed. The first approach uses a two-level learning method which consists of simple clustering of interest patches extracted from a set of training images for each posture, in addition to learning the nonparametric spatial frequency distribution of the clusters that represents one posture type. The second approach uses a two-level learning method which involves convolving interest patches with filters and in addition performing joint boosting on the spatial locations of the first level of learned parts in order to create a global set of parts that the various postures share in representation. Experimental results on video from actual US Marine training exercises are included.