Cross-Modality Person Re-Identification Based On Deep Feature

Person re-identification(Re-ID)refers to the identification of target pedestrians from existing images or videos obtained from non-overlapping surveillance cameras.Because these cameras do not overlap,there is a big difference between the field of view.In the process of re-identification,in addition to the difference in the appearance of the target pedestrian caused by different cameras,there is also the influence between different individuals.At present,most of the studies focus on visible images Re-ID,cross-modality infrared-visible Re-ID research is still in its infancy.At the same time,there is a huge difference between two modalities,which makes cross-modality Re-ID very challenging and meaningful.Thesis,based on convolutional neural network,aims at how to learn the modality sharing feature and reduce the difference between different modality,combines pedestrian attributes and auxiliary modality,and puts forward two different solutions that can effectively enhance the cross-modality Re-ID performance.The main research content is as follows:Cross-modality Re-ID based on pedestrian attribute information.Firstly,learning the thought of pedestrian attribute in the traditional Re-ID,two modal images are connected by the invariable pedestrian attribute,and the modality-invariant information is extracted from the images.The cross-modality Re-ID dataset SYSU-MM01 was manually labeled with attribute labels,and each pedestrian in the dataset was labeled with 8 pedestrian attributes.Secondly,an end-to-end neural network based on Resnet is proposed,which use pedestrian attribute labels and pedestrian identity labels for training and can extract modality-invariant local features.Through integration,the model can increase the inter-class distance while reducing the intra-class cross-modality difference.Finally,a large number of experiments on the SYSU-MM01 prove that the learned feature containing the modality-invariant information can effectively solve the cross-modal Re-ID.Cross-modality Re-ID based on mixed modality.In order to solve the modal difference between infrared and visible,an algorithm based on mixed modality is proposed.Firstly,by introducing intermediate mixed modal as the bridge between infrared and visible,the infrared and visible images are mapped to the common feature space to reduce the difference of pedestrian features,and based on this idea an infrared-mixed-visible Re-ID network is designed.Secondly,the training of network is mainly guided by the proposed multi-modality gap constraint and identity classification constraint,which can effectively reduce the difficulty of cross-modality learning.Finally,experimental results on SYSU-MM01 and Reg DB demonstrate the effectiveness of the proposed algorithm for cross-modality Re-ID.