Person Re-identification Under Intelligent Monitoring

Person re-identification refers to identifying the specific target pedestrian from a sequence ofimages or videos again by using the existed computer vision technology.The scientific research of person re-identification has high practical value.In recent years,with the development of deep-learning technology,massive growth of video image information,and rapid ascension of computer hardware performance,the main research method of pedestrian re-identification has been changed.The traditional method of feature extraction combined with distance measurement based on human design is turning to feature extraction and identification based on deep learning structures.There are two primary issues in person re-identification research at present.On the one hand,most of the existed methods of person re-identification do the pedestrian detection and person re-identification separately.It makes the pedestrian re-identification use default labeled pedestrian area.However,we usually need to extract the target pedestrian from the whole scene image,which is collected by the surveillance video in real-world scenarios.On the other hand,the quantity of pedestrian samples is too little in the common datasets,it couldn't meet the demand of training the pedestrian re-identification system.The generation of datasets for pedestrian re-identification needs a lot of preparatory work,it would need plenty of human work to complete from recording image information to labeling and classifying of pedestrians.According to the problems mentioned above,this paper has adopted and improved the pedestrian re-identification system based on the end-to-end network model.It unifies the pedestrian detection and identification into one system,connects the two parts into an organic whole by the method of sharing the pedestrians' feature information,implements the process of extracting the pedestrian from the whole image and then re-identifying the target,which would benefit its practical application more.At the same time,this paper uses the generative adversarial networks(GANs)to extend the common datasets,applies the generated pedestrian images as unlabeled pedestrian information to the training process of re-identification networks.It makes the deep neural network could learn more characteristic information and improve the accuracy and practicability of the existed person re-identification systems.Finally,this paper implements the related system according to the method and goals above.The results of the experiment have verified the system's robustness and effectiveness explicitly.