| Person re-identification technology plays an important role in the monitoring system,and its academic research and practical application have attracted much attention.Although various algorithms are constantly updated,pedestrian images collected in real scenarios have some problems,such as occlusion,illumination,low resolution,much noise,and poor discriminative ability of appearance features.As a result,existing algorithms are always not effective in person re-identification.Moreover,person re-identification is also of great significance for safety management,handling emergency incidents and cracking down on safety crimes.Based on this,this paper is devoted to the research of person re-identification.Firstly,in order to solve the problem that features in pedestrian images can not be represented clearly,this paper proposes to calculate the global relational attention of pedestrian images and supplement the global information in the network structure.Then,in order to further reduce the interference of background noise,the attitude estimation network is introduced in this paper as the pre-processing,the key features of pedestrians are extracted and the feature learning is carried out by using the graph attention mechanism,and the mutual information of channels is integrated to further enhance the representation,so as to improve the recognition accuracy of the algorithm.Finally,a person re-identification system is developed according to the proposed algorithm and applied to the real scene.Specific work contents are as follows:(1)A person re-identification algorithm based on multi-scale integrated global relation attention is proposed.Firstly,by calculating the relationship between the input image in space and channel,attention is paid to the global information on the micro level.Then,the idea of multi-scale is used in the network structure,and the output of each layer is fused to make up for this global information from the macro perspective.Finally,our algorithm is tested on four public datasets and the downsampled datasets for low resolution situation.We compare the proposed algorithm with the existing algorithm.The experimental results prove that the algorithm has superior performance and robustness.(2)A situation algorithm for graph attention based on high-order information is proposed.Based on the idea of highlighting pedestrian features in the previous algorithm,the algorithm futher optimizes the representation of pedestrian features.Firstly,the trained attitude estimation network is used to preprocess the input image to reduce the interference of background noise,and the key features of the human body are extracted as high order information.At the same time,the compressed excitation operation is used to increase the complementary information of the channel and the key point information is fused as the final representation.Secondly,the graph attention mechanism is used to learn the extracted high-level information of pedestrians.Finally,tag smoothing cross entropy loss and triplet loss are combined to conduct supervision training.Our algorithm is compared with the existing algorithm on the open data set and the low resolution data set under sampling,and the experiment proves that our algorithm has superior performance and robustness.(3)According to the proposed algorithm,a person re-identification system based on a variety of attention mechanisms is developed to solve the problem of unclear shooting in real scenes.The python programming language is used to complete the algorithm,module and framework design,and the developed person re-identification system is used to display the visual results of the proposed algorithm,and the real scene data is used to test the system.The test results show that our person re-identification system can solve practical problems and achieve the purpose of system development. |
