A Deep Pyramid Deformable Part Model For Pedestrian Detection

With the global economic picking up,the development of science of technology has also been a new breakthrough,artificial intelligence technology has also been more and more attention.Pedestrian detection technology,as an important part of artificial intelligence technology,has made great process.In daily life,pedestrian detection technology has been successfully applied in many fields,making great contributions to social security and convenience.But for the different size of people which in the complex scenes and pedestrian-intensive,the effect of pedestrian detection is still not ideal.This paper describes pedestrian detection techniques in twofold:pedestrian detection based on traditional statistical learning methods and deep learning.For pedestrian detection methods based on traditional features,it is difficult to design a feature which could express the pedestrian object well.For the traditional deep learning algorithm,the fixed size input of the network affects the detection efficiency of pedestrians with different sizes.In order to detect different size of pedestrians in intensive situations,this paper proposes the DP2MPD algorithm that combine the features extracted by the deep convolutional neural network with the traditional DPM algorithm,and contribute in two folds.First,aiming at the improve the drawback of two kinds of pooling method in convolutional neural networks,this paper proposes a novel pooling method called mixed pooling,which could select two kinds of methods by randomly number and could detect the pedestrians better during the day and night.Secondly,the pedestrian detection based on deep pyramid deformable part model,the color space pyramid is established for the different color channels of the input images,and the characteristic map of the pooling layer is extracted by deep convolutional neural network.Using the sliding window algorithm to find the position of pedestrian object in pyramid,and finally post-processing the detection frame by using the non-maximal suppression and the bounding box regression algorithm.Finally,in the experimental section,we experiment the effects of different parameters on detection results from different perspectives,and extensive experiments show that our method in complex scenarios could perform significantly better than many competitive pedestrian detection algorithms.