Research And Implementation Of Pedestrian Path Generation System Based On Person Re-ID

With the popularity of computer vision technology in recent years,a large number of related technologies have been brought into practical application,among which security applications are one of the largest application fields.At present,face recognition is still the most important technology for security applications.However,in actual application monitoring scenes,high-quality face images are not so easy to obtain due to some problems such as camera angles and lighting.At this time,it is necessary to rely on the images of human body to identify pedestrians.Therefore,this paper proposes a pedestrian path generation system based on person re-identification,hoping to implement the person re-identification technology into practical application scenarios.The research content and engineering workload of this paer mainly include the following parts:(1)Fine-grained extraction method for spatial and channel feature suppression.In this thesis,through the visualization of feature maps during the experiment,it is found that most CNNs will focus too much on important areas during the training process,and then ignore other areas of the target.Based on this phenomenon,this paper proposes a method for suppressing the regions with the strongest response in the feature map during the training process.And experiments are performed to verify the method.(2)Person re-identification method based on human 3D reconstruction constraints.In this dissertation,the person re-identification task is considered as the matching task of the same person's 3D Mesh under different cameras,different scenes and different perspectives.Therefore,this paper proposes to introduce 3D-level supervision information for person in the person re-identification data set during the training process,using multi-task learning to introduce constraints on 3D model reconstruction for the model,and to complete the missing information from the 3D to 2D.(3)Annotations for related data sets.Because the demand scenario of the cooperation unit is indoor,and the current person re-identification datasets are mostly outdoor scenes,which do not meet the requirements,this paper labels a person re-identification dataset for indoor scenes.In addition,in order to solve the problems of person incompleteness and poor lighting detected by current pedestrian detection technologies,this paper also labels a dataset of pedestrian image quality assessment to identify whether the detected pedestrians meet the requirements of pedestrian re-identification tasks.(4)Design and Implementation of a pedestrian path generation system.This system is based on the research of person re-identification related technologies in this paper.It integrates technologies such as video stream acquisition,person detection,and pedestrian image quality assessment,which solves most problems in practical applications and greatly improves the system's effect.In addition,the system also builds a related front-end interactive interface for user convenience.This article starts from the person re-identification research and explores it step by step.In the process of building the system,we find problems in the real scene and the irrationality of the design.,and we solve these problems...