The performance of an object recognition system depends on both object representation and classification algorithms. On the one hand, Object representation by using local descriptors have become a very powerful representation of images. On the other hand, SVM has shown impressive learning and recognition performances. In this paper, we present a method for fast pedestrian classification by combining a SVM with a hierarchical codebook of local features augmented with reliable global features. When compared to SVM based on local matching kernels, our method provides significant improvement of recognition performances with a speed up in learning and classification time. We evaluate our approach on a set of far-infrared images where pedestrians occur at different scales and in difficult recognition situations. The experiment shows that our method performs a fast and reliable pedestrian recognition system.