Research On Person Re-identification Technology Based On Feature Fusion

As urban population expanding,the pressure of comprehensive management of municipal safety is increasing rapidly.Therefore,the need for pedestrian detection,tracking and gender recognition is becoming significantly urgent.With the development of computer vision and deep learning technologies as well as the gradual improvement of infrastructure in the process of urbanization,person re-identification technology has become the core content of the smart city and intelligent security construction.Therefore,the development of person re-identification technology become central attention by researchers.However,there are issues in the practical application of person re-identification technology,such as poor cross-domain ability and generalization ability of the model,and poor recognition of pedestrian image blocking.Seeking the above problems,this work studies and improves the feature extraction and metric learning of the model.The main research contents are as following:Aiming at the poor generalization ability and cross-domain test performance of the traditional person re-identification model,this paper designed the Attention cascade module and the Part-level branch based on local features,and proposed the PLAN based on the fusion of Attention and local features.The attention cascade module is used to obtain pixel-level refined spatial semantic features,and the Part-level branch is used to extract multi-channel local features.After the integration of attention and local features,the model can learn refined and diversified features,thus improving the cross-domain and generalization of the model.The loss function in metric learning is not flexible enough to optimize the similarity of different samples and people are occluded,which affect the adaptability and generalization of person re-identification model.In this paper,Bach Drop module and Circle Loss function are added into the design,and a BDPLAN optimized based on Bach Drop module and Circle Loss function is proposed.Bach Drop module can make the model extract more discriminating features of self-adaptively erasing the occlusion part in the image features.Circle loss function adaptively matches the learning weight through the similarity degree within and between classes.Different similarity degree can obtain separate optimization learning weight,so as to better optimize the model and increase the adaptive ability of the model.Finally,the network model we designed is verified on the latest public data set of person re-identification,and abundant ablation experiments are carried out for separate optimized modules.Meanwhile,the network model is entirely compared with the classical person re-identification models in recent years.The person re-identification network model designed by us has achieved high recognition accuracy in both single-domain and cross-domain cases,which further prove that the introduced module designed by us has obvious validity for the generalization ability and adaptive ability of the person re-identification model.