The Research On Person Re-identification Based On Multi-attributes Fusion

Major cities in China are devoted to building the Safety City system.As an integral part of the Safety City system,person re-identification has drawn great interest in researchers due to its application and research significance.It aims at studying how to figure out a person of interest in different cameras.The visual appearance of a person is affected by a variety of factors such as resolution,illumination,viewpoint,pose and partially occlusion.The conventional hand-crafted based features have limited representation ability for large intra-class and small inter-class variations.The deep learning method have a better representation for the features,it solved the problems to a certain extent,even though,the discriminative ability still have space to improve.Based on the attribute information and deep Convolution Neural Networks(CNN),three different person re-identification algorithms are proposed.(1)Combining the classification and verification CNN methods with person ID labels,a joint learned person re-identification algorithm is proposed to.The main innovation point is the Constraint Contrast Verification Loss Function designed to impose an additional constraint to the feature values so that the metric distance can be limited in a certain range.Both the verification and classification objectives are iteratively optimized to learn more discriminative pedestrian features and measure the relationship between pedestrians in the meantime.(2)Based on the person multi-labeled attribute identification network,an algorithm that simultaneously learns multi-attributes and person IDs is designed to solve the person re-identification issue.The main innovation point is that the attribute training labels are built from automatic annotations rather than time-consuming manual annotations.The pedestrian attributes are robust to appearance changes,thus are very helpful for person re-identification.(3)Based on the above two algorithms,a multi-attributes fused solution is proposed.The major innovation point is that the proposed end-to-end framework integrates an ID verification loss,an ID classification loss,a number of attribute verification losses,and a number of attribute classification losses,and back-propagates the weighted sum of the individual losses.The combination of multiple strategies can make the various components complement each other and further improve the person re-identification accuracy.Experiments are conducted on the Market-1501 and PRW datasets.The results show that both the combination CNNs and the multi-attribute method achieve competitive performance and the multi-task method significantly improves the re-identification accuracy.On the Market-1501 dataset,the Cumulative Match Characteristic Rank-1 accuracy is improved to 70.0%,and the mean Average Precision is improved to 45.7%.The proposed method outperforms most of the state-of-the-art methods.