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