Research And Application Of Person Re-Identification Technology Based On Surveillance Video

In recent years,surveillance video networks have been widely used in various industries.The resulting massive video data cannot meet people's requirements using only traditional manual processing.Intelligent monitoring systems combined with computer vision technology have become an important way to solve this problem.As an important task in intelligent monitoring systems,pedestrian recognition has gradually attracted people's attention.Pedestrian re-identification refers to the process of matching pedestrian images under different cameras to find target pedestrians in different videos.It is of great academic significance and great application value to study its related technologies and engineering applications..In order to study the pedestrian re-recognition in actual surveillance scenes,this article starts with the original surveillance video,studies moving object detection and human detection in the video,combines feature extraction and metric learning algorithms to find the target pedestrian,and builds a pedestrian re-identification system to complete the surveillance video.Pedestrian re-identification work in progress.The main work of this article is as follows:(1)Aiming at the problems of high computational complexity or inadequate appearance description of traditional Color Naming(CN)features,a relatively simpled CN feature is proposed,and a new nonlinear strategy is adopted to Hue,Saturation,Value(HSV)color space,quantizing the color space to 10 levels,using low-dimensional vector space to represent highdimensional color space,and better describing color characteristics.(2)Research and implement the pedestrian detection algorithm in the video.Firstly,the moving object detection algorithm is researched,and the Vibe algorithm is selected for detection.The moving target area is obtained,and the pedestrian precise inspection is performed in the moving area,which greatly improves the pedestrian detection efficiency.Secondly,the improved deformable component model is used for pedestrian precise inspection.For the shortcomings of the Histogram of Oriented Gradient(HOG)feature in describing the target,the simplified HOG feature and Local Binary Pattern(LBP)features.At the same time,an adaptive weight allocation method is proposed to assign different weights to different parts of the model,so that different parts are more distinguishable and effectively solve the problem of pedestrian occlusion.Finally,for the problem of high complexity of feature pyramids in the model,the construction process of image pyramids was optimized and the algorithm complexity was reduced.Through experimental comparison,the superiority of the algorithm is verified.(3)Research and implement the pedestrian re-identification algorithm.First of all,in order to improve the detection efficiency,combined with the designed CN color features,the pedestrians are pre-recognized by comparing the color of the upper and lower half of the pedestrian to reduce the range of the target to be detected.Secondly,in order to solve the problem that the traditional simple low-level feature fusion cannot solve the target image's low recognition rate due to illumination and scale changes,this paper combines the intermediate CN feature and the simple high-dimensional low-level feature to obtain a more robust fusion.Pedestrian characteristics.Finally,the most appropriate metric learning algorithm was selected through experiments and tested on multiple data sets,which verified the effectiveness and applicability of the pedestrian re-identification algorithm.Finally,based on MFC and Opencv open source libraries,a video-oriented pedestrian reidentification system is designed and built.The video pedestrian detection and pedestrian reidentification module are connected in series to complete the image and simple semantic pedestrian re-identification work.