Sparse shift-invariant representation of local 2D patterns and sequence learning for human action recognition.