Traditional object detection method needs to judge a large number of candidate windows (regions),so it needs a large amount of calculation.In this pa per,according to pedestrian characteristics,we put forward a method based on hierarchical judgment.T he candidate windows that need to be detected is reduced progressively,so we can reduce a large number of can didate regions that need to be judged by complicated feature and classifier,which reduces the amoun t of calculation of the whole algorithm.Firstly,we extract norm of the gradients (NG) feature from image,and use linear support vector machines (SVM) to get the candidate regions of the ped estrians.Secondly,we extract simple Histograms of oriented gradients (HOG) feat ure from the candidate regions,and the candidate regions are further filtered by linear SVM.Finally,we extract multi-resolution HOG feature from the filtered c andidate regions,deformation part model (DPM) to detect the candidate areas to locate the precise location of pedestrians.On the INRIA data set,experimental results show that on the basic of ensuring the accuracy of det ection,a small number of pedestrian are not detected compared with the original DPM algorithm,but the number of err or detection is far less than that of the original DPM algorithm,and the detection speed is faster than that of the origin al DPM algorith m.