Sparse coding using learned dictionaries can adaptively represent signals. However

the similarity among the signals that are encoded in traditional dictionary are lost due to a lack of correlations between atoms. Considering the robustness and discriminative power of structured sparse representation

the building of the structured dictionary becomes an important task. We conceive a framework of tree-structured dictionary by introducing a constraint for the data point code path (programming the index from the upper layer to the next layer) according to the standard convex optimization dictionary-learning algorithm. Experimental results on the KTH human action database show that local descriptor codes with learned tree-structured dictionary have good robustness and discriminative and demonstrate that our algorithm generally obtains higher recognition accuracy than other similar methods. We achieve an accuracy rate of 97.99% using histograms of oriented 3D spatial-temporal gradients(HOG3D). From our experiments

we observe that the encoding of signals using the constructed dictionary has good robustness and discrimination

and is preferable for the task of classification.