Pedestrian Detection Based On Deformable Part Model Under Haze Weather

The rise of computer science has promoted the theoretical research and practical application of pattern recognition and machine learning.Pedestrian detection is the focus of research in this field.The complex detection environment and its large non-rigid characteristics bring great challenges to its wide application.Compared with pedestrian detection under normal illumination environment,the performance of existing algorithms is seriously degraded under complex conditions,such as diverse background,foreground occlusion and haze,which are greatly affected by illumination.If target detection is carried out directly on images collected under complex conditions such as haze weather,there will be serious false detection and missed detection.Therefore,in order to solve this problem,a pedestrian detection algorithm based on Deformable Part Model(DPM)in haze weather is proposed to improve the efficiency of pedestrian detection in haze weather.The main research contents are as follows:Firstly,it is necessary to identify whether the input image is haze or not.Through a comprehensive comparison of a large number of clear images and haze images in the same environment,it is concluded that there are obvious differences in mean square error,peak signal-to-noise ratio,structural similarity and information entropy between them.In this paper,image information entropy is used as the quality evaluation index.We compare the information entropy of the input image with the set threshold,and determine whether the operation of defogging is needed.Secondly,we need to defog the haze pictures.At present,the technology of image defog mainly includes image restoration technology based on mathematical modeling and image enhancement technology based on human vision.By comparing the defogging effect of the classic defogging algorithms such as Retinex and Dark Channel Prior(DCP),this paper uses DCP as the basis to improve the global illumination intensity estimation process in the algorithm,and solves the color spots and distortion in the defogging process.At the same time,we indirectly improve the calculation process of transmittance by downsampling the input image,so as to improve the efficiency of fog removal.The improved defogging algorithmlays the foundation for the subsequent pedestrian detection and location under the haze weather.Finally,pedestrian detection algorithm in haze environment was proposed which is based on DPM.Comparing the detection effects of HOG+SVM,Haar+Adaboost,DPM and other classical pedestrian detection algorithms on haze pedestrian data sets.In view of the drawback of DPM detection speed,this paper reduces the number of feature maps and candidate windows in the target detection process by using the regional proposal method instead of the traditional sliding window based on the DPM algorithm.And it is combined with the improved DCP algorithm to obtain a solution that can improve the detection precision while improving the detection speed under the haze weather.The experimental results show that the pedestrian detection accuracy of the proposed algorithm in haze environment has been improved from 76.64% to 89.94%,which basically meets the needs of practical application.