Due to the resolution of small size pedestrian is relatively low, and the hard negative background is very similar to people, therefore, detecting small size pedestrian or detecting pedestrian from hard negative background still a challenging problem in computer vision. In order to effectively address these problem, we propose a novel deep convolution neural networks, and cascade an improved boosted forest classifier method to detect pedestrian. Firstly, by using selective search method to propose pedestrian candidate boxes with confidence scores for utmost retaining image resolution; then, based on these proposed confidence values, adopting convolution neural network model to extract candidate regions feature maps; finally, we improve the boosted forest classifier and cascade it to classify candidate boxes for achieving efficiently pedestrian detection. Extensive experiments on Caltech and KITTI benchmarks demonstrate the proposed method outperforms the state-of-the-art, achieves promising precision on KITTI and the lowest miss rate of 11.53% on Caltech, outperforming the second best method (CompACT-Deep) by 0.17%.