Research On Person Re-identification Based On Deep Learning

With the development of image and video recognition technology,and the global attention on urban security systems,pedestrian re-identification has become a research hot spot in many fields such as surveillance video intelligent analysis,personal album management,large-scale pedestrian flow analysis.It has a wide range of application scenarios and will play a huge role in maintaining social security and improve the efficiency of criminal case investigation.Person re-identification is essentially to solve the pedestrian image matching problem across scenes.The locked target disappears from a camera.When the target appears in other camera perspectives,the system can still re-lock it according to its characteristics.At present,the difficulties in solving this problem mainly include the interference of illumination changes,the diversity of pedestrian attitudes,and the lack of time and space information.Therefore,person re-identification is also a challenging research task.This paper employs research methods of deep learning to improve the matching accuracy of person re-identification and reduce the workload of sample labeling.Aiming at the problem of low recognition and poor stability of pedestrian image extracted from early person re-identification research,this paper improves and further studies person re-identification based on using the convolutional neural network the extraction of pedestrian features,and proposes the person re-id research methods based on multi-layer depth feature fusion.The pedestrian image is processed by a classical convolutional neural network Le Net-5.The features obtained by each layer of the convolutional neural network are subjected to PCA dimensionality reduction,retaining its main components,and each the characteristics of the layer after dimension reduction are merged.The similarity between the pedestrians to be inquired and the pedestrians in the image library is determined based on the Euclidean distance,and the re-recognition result is obtained.The experimental results show that compared with the existing personre-identification method,the accuracy of this method is higher.Aiming at the problem that the current person re-identification cross-data set test performance is seriously degraded and the training sample labeling cost is high,a method of adopted style migration and metric fusion is proposed.The cycle generative adversarial networks is used to convert the data image style of a data set in one data set to another unlabeled data set.Training on image data after style conversion,and using a combination of direct metrics and indirect metrics for similarity metrics.Testing on unlabeled datasets and arranging pedestrian images from high to low.The experimental results show that the recognition performance of person re-identification in across data sets can be significantly improved by this method.