Person Re-identification Based On Convolutional Neural Networks

Due to the rapid development of video surveillance,the massive video data has brought great challenges to the traditional artificial video analysis methods.Person re-identification(Re-ID)is to identify the same person's images coming from non-overlapping camera views.Images for re-identification are mainly captured by different cameras.Person re-identification has drawn intensive attention in the computer vision for a decade.It is a critical task in video analysis,surveillance,human-machine interaction and so on.However,it is still an unsolved problem due to the dramatic variations caused by illumination,view point,pose,occlusion and resolution.There are two main components in person re-identification,feature extraction and metric learning.The traditional person re-identification methods rely on the hand-crafted pedestrian features.However,since the same pedestrian may be very different in different images,and different pedestrians may look very similar,it is difficult to apply these manual features to the complicated real environment.Convolutional neural networks(CNNs)have achieved great success in many computer vision tasks,including image classification,object detection and face recognition.They are also applied to person re-identification.But the problem is still far away to be solved.In this paper,CNNs are applied to the study of Person re-identification.The research contents include:1?Person re-identification based on Classification model.Different from the commonly used contrastive loss and triplet loss for similarity measure,we train a CNN using SoftMax loss for person classification.Firstly,by fine-tuning the pre-trained AlexNet using Pedestrian dataset,it is used for feature extraction,replacing the hand-crafted features in traditional methods.Secondly,Design a specific CNN for person classification.Then extract convolutional features using the CNN for person Re-identification.2?Design a model with improved Siamese structure for person re-identification based on CNNs.We design a CNN architecture and a loss function composed of SoftMax loss and contrastive loss for feature extraction,and then use metric learning algorithms for re-identification.Experiment results on two public datasets demonstrate the effectiveness of our model.The Rankl index on VIPeR database and iLIDS dataset are 41.8%,and 54.3%respectively,outperforming most of existing models.