The multi-scale orientation (MSO) features for pedestrian detection in still images are put forwarded in this paper. Extracted on randomly sampled square image blocks (units), MSO features are made up of coarse and fine features, which are calculated with a unit gradient and the Gabor wavelet magnitudes respectively. Greedy methods are employed respectively to select the features. Furthermore, the selected features are inputted into a cascade classifier with Adaboost and SVM for classification. In addition, the spatial location of MSO units can be shifted, are used to the handle multi-view problem and assembled; therefore, the occluded features are completed with average features of training positives, given an occlusion model, which enable the proposed approach to work in crowd scenes. Experimental results on INRIA testset and SDL multi-view testset report the state-of-arts results on INRIA include it is 12.4 times the faster than SVM+HOG method.