The Impact Of DPM Model Parameters On Pedestrian Detection Results

Based on the pedestrian detection results,the influence of the Deformable Parts Model's parameters were studied in this paper by using different parameter combinations trained DPM pedestrian models.The optimal DPM model in pedestrian detection was got by the analysis of the evaluation indexes as the true positive rate,the false positive rate and the false negative rate.HOG feature was extracted from 4 to 14 cell size and the DPM pedestrian models trained with 6 to 10 parts in the INRIA Person annotation image library.The pedestrian detection experiments were performed in a set of 250 positive and negative images randomly selected from the INRIA Person image library,and in the MIT pedestrian library with 924 original images.Analyzing the detection results,a good pedestrian detection model was achieved by using DPM models trained in one component,with 8 or 9 parts,as sbin is 6 or 8.The results may provide theoretical support on the parameters selection in DPM training.The robustness of the selected optimal parameters DPM model was verified on the MIT pedestrian image library and the self-built image set of 231 images in natural environment.The models performed excellent in the given pedestrian image libraries which can balance different evaluation indexes.The true positive rate reached from 0.844 to 0.918 and the false detection rate among 0.039 to 0.097.Above all,excellent target detection effects and good robustness can got by using "6*6" or "8*8" HOG features,the DPM pedestrian model trained with 8 or 9 parts in 1 component.The results can give us a good tradeoff among various evaluation indexes in pedestrian detection.