Research On Pedestrian Re-identification Algorithm With Improved Representation And Metric Learning

As the most important branch of artificial intelligence,a number of practical and fast-growing applications have derived from computer vision.Due to the large deployment of intelligent video surveillance systems in recent years,intensive surveillance networks have been formed in many public places,making it difficult to rely solely on manpower to observe pedestrian targets and to track them across cameras.Pedestrian re-identification has gradually become a hot topic in computer vision.It mainly studies how to retrieve and identify key pedestrian targets in the multi-scene multi-camera surveillance system.Its essence is to match multiple pictures of the same pedestrian in different cameras.Because of the varieties in lighting conditions,pedestrian attitudes,observation angles,etc.in actual application scenarios,multiple images of the same pedestrian tend to have large differences in appearance,thereby bringing great challenges to matching them correctly.This paper proposes a pedestrian re-recognition algorithm that improves representation learning and metric learning.It studies three factors that influence the validity of pedestrian re-identification: feature representation,similarity measure and reranking of matching lists,and then improve and optimize the re-id algorithm.Firstly,in the feature extraction stage,the high-precision classification network DenseNet is used as the feature extraction network to extract more robust and reliable pedestrian represent features.Then the singular value decomposition(SVD)is used to reduce the correlation between similar pedestrian represent features of the base network output,to reduce the error of similarity calculation and to optimize the deep learning process.Using the improved Softmax loss function in the metric learning process,which makes the pedestrian features extracted by the network have larger inter-class distance and smaller intra-class distance;In this paper,we also use the reranking method based on k-reciprocal nearest neighbors to rerank the pedestrian images that need to be retrieved in the gallery,and further improve the accuracy of re-id.The rank-1 accuracy on the test data sets of Market1501 and CUHK03 reached 88.32% and 59.1%,respectively.This indicates that the method proposed in this paper can significantly improve the accuracy of pedestrian re-identification and reach the state of art.At the same time,the method also showed excellent results in the test of the actual monitoring system.