We propose a new pretreatment for pedestrian detection with convolutional networks. It is widely known that the phenomenon of overlapping feature distribution is common, which leads to overfitting problem. We present a method that divide one category that have overlapping distributed features into multi-subcategories. By this means smooth boundaries can be easily found to separate different subcategories, and overfitting can be avoid as well. And during the detection process, we added a fusion layer before the classifier, in the fusion stage we convert the multi-subcategories classification back to the original classification problem. The method we proposed here has been thoroughly tested with over 2,000 pedestrian images in ImageNet, and we have achieved a high accuracy at 68.1%.