Person Re-identification Research Based On Mask RCNN Neural Network

With the rapid development of information technology and the increasing emphasis on urban security,person re-identification,as one of the key technologies of intelligent video security system,is receiving more and more attention.At present,How to get more powerful discrimination and generalization ability of pedestrian features and how to apply person re-identification together with pedestrian detection to real scene are two important directions of person re-identification research.Acording to the study of person re-identification,it is found that the traditional person re-identification datasets are obtained by pedestrian detection.The images always contain many interferences such as background,occlusion,other pedestrians etc,which affect the effect of pedestrian feature extraction.In order to deal with the problem,this paper uses Mask RCNN instance segmentation neural network to remove the background etc interferences while detecting pedestrians.Considering that the objects carried by the pedestrian are helpful for the representation of his features,a method of person re-identification feature extraction combined with these auxiliary objects is proposed.The image features are not used enough in the image segmentation part of Mask RCNN,which leads to the poor effect of object segmentation,a reference of semantic segmentation method is used to improve this part,and completed the construction of the improved Mask RCNN instance segmentation neural network.At the same time of pedestrian instance segmentation,auxiliary objects such as backpacks,umbrellas,handbags and suitcases were reserved,combined with pedestrian features to get pedestrians'mixed features at person re-identification stage,and a person re-identification neural network model is designed.At the end of the paper,a person re-ID neural network model that can be used to real scene is constructed.The validity of the improved methods of Mask RCNN and person re-identification neural network model are verified by experiments on public datasets.The experimental results show that the improved methods of Mask RCNN can make the result of instance segmentation better,the person re-identification neural network can effectively identify the targets in the datasets and improve recognition accuracy.Therefore,the research has certain theoretical significance and application value.