Research On Person Re-identification With Local Features Fusion Based On Deep Learning

Person re-identification(Person Re-ID)technology is an important foundation for intelligent security and target tracking,and plays a significant role in safe and smart city.Due to the differences in the deployment positions of surveillance cameras,some problems in pedestrian images such as different pedestrian posture,variation of illumination and diverse shooting angles are presented.These difficulties result in that less person information are available,which make Person Re-ID still a very challenging task.Therefore,how to design an algorithm that captures and organizes the local features of images efficiently,and improve the accuracy of Person Re-ID network,has important research value.Recently,due to the strong robustness for Image occlusion and posture change,the PCB(Part-based Convolutional Baseline)network has drawn considerable attention.However,the accuracy of the algorithm in Person Re-ID task is still low,which means it has a lot of room for improvement.This paper improves this algorithm from different aspects,and the main research contents are as follows.(1)Most of the local features based Person Re-ID methods met the problem of the lack of robustness and discriminability due to the distortion of information in the procedure of extraction of pedestrian features.A novel Re-ID algorithm based on multi-level feature fusion with overlapping stripes is proposed.Firstly,the features with overlapping stripes are extracted to compensate the loss of information.Secondly,the hierarchical feature fusion strategy is adopted,which integrates the deep and shallow features of the target,to make the network model distinguish the detailed information more effectively.In addition,group normalization modules are designed to reduce the optimization differences within various loss functions for obtaining appropriate shared features.Finally,the experimental results on the Market-1501 and Duke MTMC-re ID datasets show that the proposed algorithm can make full use of the person features,and improve the accuracy.(2)Most of the multi-branch network based Person Re-ID methods met the problem of the lack of the heterogeneous features in the procedure of extraction of pedestrian features.A novel Re-ID algorithm based on heterogeneous branch correlative features fusion is proposed.Firstly,the attention-based OSNet is designed as the backbone sharing network,which increases the weight gap of different channels and extracts more significant and distinguished key features.Secondly,the stripe features correlation module is developed,which compensates the loss of information while enhances the relatedness of the parts of the body.In addition,the heterogeneous features extraction module is designed to increase the structural diversity of the model for learning difference features.The single ID prediction loss is proposed to improve the classification and prediction ability of the model.Finally,the experimental results on the Market-1501 and Duke MTMC-re ID datasets show that the proposed algorithm can effectively reduce the interference of irrelevant features,and improve the discriminative ability of features.(3)Most of the spatial partition based Person Re-ID methods met the problem of the lack of the longitudinal semantic information in the procedure of extraction of pedestrian features.A novel Re-ID algorithm based on multi-partition features fusion is introduced.Firstly,the highly active drop module is proposed,which create a mask to cover the most activated regions.The learning ability of the model for the features with high discrimination in activated regions is improved.Secondly,the pyramid spatial partition module is designed to integrate the local features of different granularity,which helps to learn the person local features efficiently while keep the completeness.In addition,the channel partition module is developed and complementary with pyramid spatial partition module,which highlights the importance of semantic concepts and improves the learning ability for the features with a wealth of semantic information in local regions.Finally,the proposed algorithm is experimentally verified on the Market-1501 and Duke MTMC-re ID datasets,and the results show that the algorithm could extract the more detailed features,and reduce the false matching rate efficiently.