Research And Application Of Person Re-Identification Technology Based On Deep Learning

Person Re-Identification is one of the important directions in the field of computer vision research,which is of great significance to the maintenance of social security and the construction of a harmonious society.With the development of deep learning,many scholars have applied deep learning technology to Person Re-Identification and achieved high accuracy on the corresponding datasets.However,due to the different illumination levels of day and night,and the extensive use of infrared cameras in real scenes,the matching problem of pedestrians in color modality images and infrared modality images appears.Therefore,cross-modality Person Re-Identification is facing a huge challenge.Based on deep learning technology,this thesis intensive studies crossmodality Person Re-Identification,which is one of the hot research directions of Person Re-Identification in recent years.The main content and innovation points of this thesis are summarized as the following aspects.Firstly,this thesis studies cross-modality Person Re-Identification in depth from the aspects of definition and research difficulties.In view of the difference between the color image and the infrared image in the two modalities,as well as the change of the shooting perspective and the change of pedestrian posture in the same modality,a Squeeze-and-Excitation dual-path network is proposed to learn the discriminative features of pedestrian images in two modalities to reduce the difference between two modalities.In this thesis,attention mechanism in Person Re-Identification technology is combined with cross-modality Person Re-Identification,and the difference between RGB image and IR image is analyzed from the perspective of channel domain.The Squeeze-and-Excitation module is integrated into classical dual path network,learning the features of RGB image branch and IR image branch respectively.The two subspace are mapped to the same feature space,in this way,the differences of features learned between inter-modality and intra-modality can be reduced.Then,the cross-modality hard triplets loss function and ID loss function were used for metric learning,so that the feature distance of the same pedestrian in the same modality or cross-modality increased,and the feature distance of different pedestrian decreased,which promoted model training.The proposed method is tested on relevant datasets to verify the effectiveness of the algorithm.Finally,this thesis design and realize a Person Re-Identification system,it includes the cross-modality Person Re-Identification,the cross-modality algorithm based on attention mechanism proposed in this thesis is applied to the recognition module,and the matching of visible image and infrared image is added.A complete Person ReIdentification system is constructed by combining the login module and pedestrian retrieval library building module of the whole system.Through system tests,it is proved to be effective.