Research On DPM Fast Pedestrian Detection Method Based On Pyramid Layer Location

With the rapid development of China's economy,car ownership is growing continually.The growing road traffic conditions have provided great convenience to people's life.However,the increasing number of cars leads to frequent traffic accidents,traffic safety situation is deteriorating.Pedestrians are the main participants of the traffic system,but also the main victims of traffic accidents.Therefore,the research of pedestrian detection system is of great significance.In the present study of various types of pedestrian detection systems,the pedestrian detection method based on machine learning is a hot research topic.Deformable part model(DPM)is one of the most popular pedestrian detection algorithms in recent years.The detection algorithm based on DPM splits the target pedestrian into several parts,and detects a pedestrian by detecting each part.While learning the pedestrian detection based on DPM,we have found that the high detection accuracy of DPM is at the expense of great computation burden,thus limiting its development and application.This paper is focused on solving the problem of the low detection speed of DPM.We present a fast DPM pedestrian detection method based on Pyramid level locating.Firstly,the object proposal instead of sliding window method is used to obtain pedestrian proposal regions,and the candidate windows are extracted according to the size and aspect ratio.A large number of non-pedestrian areas are excluded to reduce the number of windows which need to be calculated.Secondly,a pyramid level locating method is proposed to select the optimal root level in feature pyramid for each proposal window.Thus,we can construct a simple feature pyramid that contains a small number of feature maps.Then,a single-point calculation method is designed replacing the traditional dynamic programming method to calculate scores of proposal windows efficiently.Finally,pedestrians can be located from images.PLL-DPM method can reduce the number of feature maps and windows that need to be calculated in detection process obviously.The experimental results on the public pedestrian data sets show that it improves the detection speed effectively while ensuring the detection rate,and reduces the false detections in a certain extent.