Abstract:

Aiming at pedestrian detection problem in complex environments, we propose an effective image representation method based on hierarchical sparse coding. First, we obtain the sparse representation by a two-layer sparse coding model combined with a K-SVD based deep learning algorithm. We then model image blocks and higher-order dependencies of the same region, forming an effective unsupervised feature learning method. After that, we fuse the sparse representation with the sparse representation of the SIFT descriptor, obtaining a more comprehensive and more discriminant image representation. Finally, together with the SVM classifier, it is applied to pedestrian classification tasks. Experimental results show that the pedestrian classification method has very competitive performance in comparison with  other similar methods.

Key words: pedestrian classification, sparse coding, spatial pyramid matching, feature fusion, K-SVD