In pedestrian classification, there are many factors, such as light changes, posture changes and occlusion problems etc, which brings many difficulties for feature extraction process. A hierarchical feature method is put forward based on sparse coding. The method trains optimal sparse coding with forward prediction function, and then learns the two levels networks one by one in unsupervised manner with Convolution Predictive Sparse Decomposition algorithm (CPSD) under framework of the deep convolution network model. Then we make the feature fusion. Finally, we implement classification with SVM algorithm. Experimental results demonstrate the effectiveness of our method for pedestrian classification, which has significant performance improvement compared with similar methods.