Target Detection And Recognition Of Vehicles And Pedestrians Based On Improved DPM

Object Detection is one of the important research in the field of computer vision.Detection technology effectively helps people solve many problems,and it has been widely used in real life.Deformable Part Model(DPM)is a popular method of Object Detection technology.DPM is a object detection model based on HOG features,compares with other object detection technologies,DPM can well reduce the influence of illumination factors on detection results with a significant advantages in fuzzy visibility,moreover,it has very high recognition accuracy.But DPM still has some drawbacks.DPM has made some improvement on HOG.However,there are still some defects in the original HOG feature,which makes object detection effect of DPM not ideal.In this paper,two problems of DPM are improved.First,DPM improves HOG features and then extracts features.One of the advantages of HOG is that HOG represents edge(gradient)structural features that can describe local shape information.However,in the smooth area,HOG cannot effectively reflect the regional feature information.And when the difference between the target object and its picture background color is not obvious,the HOG cannot clearly show the edge profile between the target object and the background.Second,DPM uses scoring to identify whether there is a target detection object in the current detection picture.The score is the sum of the two partial scores: one is the root filter score,and the other is the component filters score.When there are occlusion problems with the object in the detected image,the feature information obtained by the component filter during feature extraction is incomplete,which may cause this phenomenon that when the position of the occlusion component is detected correctly,the component filter scoring is lower,so that the highest Points are reduced,which resulting in missed tests when the target is detected.In this paper,two problems of DPM are optimized,and a weighted DPM algorithm based on multi-fuzzy features is proposed.The main improvements are two points: First,a HL fusion feature based on HOG and rotation-invariant equivalent LBP features is used to get DPM feature extraction,the HL Fusion feature can not only get the gradient information,but also can effectively represent the image texture information,effectively solves the problem that the hog can not embody its characteristic information effectively in the smooth region,and also solves when the target object and its picture background color are not big,Hog is not a good representation of the edge contour between the target object and the background..Second,the weighted component model is proposed to improve the efficiency of the effective information component filter,in the process of target detection,when there is occlusion,the importance of different parts filter can be distinguished,which makes the weights of the feature information components become larger,the response points are higher,the feature information is less or the proportion of the feature information is smaller.,this can fully play the role of useful parts,avoid the loss of detection information caused by occlusion,greatly improve the detection accuracy.Experiments show that this model can effectively solve the problems of the original DPM presented in this paper and improve the accuracy of object detection.