Person Re-identification Based On Fusion Feature And Siamese Network

Person re-identification aims to match the camera images which contain the same person in surveillance videos.The difficulties of person re-identification are including different view condition,large changes in illumination,complex background environments and low image resolution.Firstly,this paper studies the attribute-based fusion feature,making the pedestrian image description more discriminative.In addition,cause of the low resolution or multi-views,the recognition of some datasets is very low.Therefore this paper studies the work of transfer learning.Considering the single network is too simple and the image description is not sufficient.In this paper,we propose a fusion model based on ResNet50 and VGG16 that used as the sub-network of Siamese network.The main contributions of this paper are as follows:1.Attribute-Based Fused Feature for Person Re-identification.In this paper,we propose an attribute-based fused feature which combined low-level and attribute features for person reidentification.Firstly,each of the image is divided into four parts and the gradient-LOMO(Local Maximal Occurrence)features are extracted separately.The sub-classifiers are respectively trained and fused to generate a classifier,which describes 21-dimensional attributes and are combined with the correction mechanism.Then,the time complexity of manual marking can be reduced effectively by using this classifier to predict attributes on unmarked datasets.Finally,the low-level features,which extracted from original and foreground images,are combined with attributes as the fused feature.The results of experiments on VIPeR,PRID450 S and GRID indicate that our method exhibits prominently performance which outperforms the state-of-the-art methods for person re-identification.2.Person re-identification based on fusion model and Siamese Network.During training procedure,a fusion model named F-CNN that based on ResNet50 network and VGG16 network is proposed and then used as the subnet of Siamese network.The subnet of Siamese network share the weights.Two loss function named verification and identification that based on softmax are considered in this part.In the testing stage,the unsupervised transfer learning model is used to transfer the model to the unlabeled dataset.Also,the similarity of the two randomly input images and the identity label of each are predicted.GRID and Market-1501 datasets are both used as the target datasets.Finally,the experiments on datasets of CUHK01,VIPeR,GRID,Market-1501 and DukeMTMC have achieved good results.