Research On Person Re-identification In Traffic Environment

Person re-identification(Person Re-ID)is to match individual images of the same person captured by different non-overlapping cameras,which are extensively deployed in every cities as parts of the urban road surveillance system.It is an important part of a smart city.In this work,we formulate a new person re-identification framework including a metric learned residual network feature extraction procedure followed by a query re-ranking.A residual network with pre-trained weights and revised CenterLoss layer is fine tuned on the Re-ID datasets,then we drop the full connection layers and use the rest of the network to extract person representation features.Compared with the traditional contrastive loss and triplet lossthe metric network based method is more robust and accurate when tackling the person Re-ID problems.The extracted feature by our network is superior to all hand-crafted features and deep learned features so far.We also conducted comparative experiments to validate the superiority and advantages of our proposed framework over state-of-the-art competitors.The results show that the framework can achieve the better performance in Rank-1 and mAP against baseline method.In addition,the training and implementation of the method in this paper is simpler,so it has better practical application value.The main work of this article is as follows:1)This paper surveys a broad selection of the hand-crafted systems and the large-scale methods in both image-based and video-based Re-ID systems.This paper also analyzed the loss layer's influence to the accuracy of the person re-identification system.2)A residual network feature extraction method based on CenterLoss is proposed.This method is superior to all other image representation methods in the field of personal identity authentication.Our algorithm is easy to train and requires little manual intervention.3)An effective and unified pedestrian recognition algorithm is constructed,and high accuracy is achieved on multiple standard pedestrian data sets.