Research Of Person Re-Identification Based On Video Object Tracking

Massive monitoring videos collected in the background of big data,providing a reliable source of data for humans to track target pedestrians.However,it is time-consuming and laborious to investigate and track targets with artificial eyes.With the rapid development of artificial intelligence technology now,we need an auxiliary system to help people free themselves from the tedious work of watching,finding,and analyzing the trajectory of target personnel.In this thesis,a pedestrian re-identification method based on video object tracking is proposed,which can be applied in daily life scenarios.Firstly,an object detection model for pedestrian detection is trained,which can obtain accurate and suitable pedestrian pictures in the video.Then,the person re-identification model with excellent performance in the cross-domain environment is trained to adapt to the use environment of style changes.Subsequently,the stability of object tracking algorithm is utilized to assist in obtaining continuous and stable tracking results.Finally,a key personnel tracking system is designed and implemented.The specific work of this thesis is as follows:1)The original YOLOv5 algorithm is mostly used to detect common objects in daily life,but its detection ability for pedestrians in complex environments is slightly insufficient.This thesis modifies the Backbone network used for feature extraction in the network structure,adds an attention mechanism to it,uses SIOU loss function to replace the original CIOU loss function,and on this basis uses Wider Person data set to train the object detection model dedicated to pedestrian detection.Finally get excellent detection ability for crowded pedestrians and small target pedestrians in complex scenes;2)The existing pedestrian re-identification model needs to strengthen its cross-domain adaptability and improve its accuracy in practical application.This thesis uses Res Ne St as the backbone network to research and design a new pedestrian re-identification model,which not only uses the channel attention of feature map,but also uses the grouping module architecture to accelerate the computational inference speed.Subsequently,various training strategies were referred to improve the accuracy of the model,and draw lessons from IBN normalization method improve the cross-domain adaptability of the model;3)The pedestrian re-identification model has high accuracy in determining the tracking target,but it is prone to miss detection in the face of the change of feature caused by the change of target’s posture in the video.In this thesis,Deep SORT,a multi-object tracking algorithm combining spatial information and time information,is used to bind the target person ID,and the continuity of the results in the object tracking algorithm is used to expand the query library and make up for the missing query track;4)In crowded scenes,pedestrian tracking algorithm is prone to ID confusion in the face of moving targets with drastic changes in motion status,and the previously bound ID may be invalid at present or trace to low-quality pedestrian screenshots.In order to avoid blindly expanding the query database and adding dirty and incorrect samples to disturb the tracking results,this thesis proposes a dual check method,which checks the stability during ID binding to ensure that the correct information is targeted.New images and new features are added to the query target database after subsequent frame checks to ensure that the target ID is not disordered and that the addition is not a mutation feature.