Person Re-identification Based On Deep Learning

Surveillance system has become more and more important in people's daily life. However,a large number of surveillance video makes manual identification a time-consuming and laborious task. Today, the demand for intelligent surveillance has been growing with time, automatic locating and recognizing the monitoring target has attracted the attention of many researchers. Person re-identification has been widely concerned due to its huge challenge.Person re-identification tells whether this person has been observed in another place (time) by another camera. Face a series of problems such as feature extraction, representation, similarity measure and so on. In this paper,we study the above problems by using deep learning methods in person re-identification. The main work is as follows,First of all,this paper proposes a person re-identification algorithm based on the Bag-of-Words model and deep features. We use the output of convolution layer as the local descriptor of pedestrian images to replace the hand-craft features such as SIFT and SURF in the Bag-of-Words model. The contrastive experiments on the Market-1501 dataset show that the proposed local feature in this paper is a more stable feature representation than the hand-craft feature.Secondly, we propose a person re-identification algorithm based on improved triplet loss function. In this paper, we construct a new TripletNet loss function. We add the constraint of the distance of the positive image pairs on the constraint of difference between the distance of the positive image pairs and the distance of the negative image pairs. The proposed method achieves a better performance on the VIPeR and CUHK01 dataset and the generalization ability of the model is also stronger.