Research On Person Reidentification Technology Of Cross-camera Under Surveillance Video

As one of the typical applications of computer vision,video surveillance has been widely used in road safety,public security,intelligent transportation and other civil and military fields.Cross camera person re-identification under surveillance video is of great significance to urban safety.Person re-identification means that for a pedestrian appearing in a surveillance video,the image of that person is retrieved across the device to determine whether it appears in other cameras.In order to realize person re-identification in surveillance video,this paper mainly completes the following work:First of all,the popular pedestrian detection algorithms based on deep learning is compared,and an improved SSD algorithm is proposed.Res Net is used as the basic network to detect multi-scale pedestrian targets,which improves the accuracy of target detection.In view of the characteristics of pedestrian detection in surveillance video,the length width ratio of the default box is improved,so that the accuracy of target location is higher.Experimental results show that the improved SSD algorithm can not only achieve faster speed,but also significantly improve the accuracy.In order to improve the recognition ability of the network to the pedestrian features,the classification model and the recognition model are combined to learn the identification features and the similarity measurement,while making full use of the re-ID annotation,using their complementarity to improve the recognition ability.This network not only has a good effect on the public data sets such as Market-1501 and CUHK03,but also has been verified on the campus monitoring video data set.Then combined with the improved pedestrian detection algorithm,the function of real-time searching specific pedestrian in surveillance video is realized.In view of the lack of samples in the existing data set for person re-identification,the deep convolutional generative adversarial networks are used to generate unlabeled samples,the CNN network is used for feature learning,and the outlier label smoothing regularization is used to evenly distribute unlabeled images.Experiments show that the data generated by GAN can effectively improve the discrimination ability of CNN,and improve other basic network models on the public data sets and the campus monitoring video data set.At the same time,the improved algorithm is combined with pedestrian detection to realize the function of real-time retrieval of specific pedestrian in surveillance video,which verifies the effectiveness and practicability of this method.