Person Re-identification Algorithm Based On Posture Alignment

Person Re-Identification refers to pedestrian recognition and retrieval in cross-camera scenarios,and is one of the hot research areas in computer vision.In recent years,researchers have made full use of computer vision and deep learning technologies to greatly improve the performance of person re-identification.However,in practical application scenarios,the study of person re-identification is still very challenging due to changes in perspective,occlusion problems,and different attitudes.Therefore,how to obtain discriminative features with high accuracy and robustness in cross-view scenes is the research focus of person re-identification.Based on the research of person re-identification algorithm based on pose alignment,this paper focuses on the improvement of pedestrian feature extraction in deep convolutional networks in person re-identification algorithm,and proposes person re-identification based on attention mechanism and.The specific work as follows:(1)Person Re-Identification based on attention mechanism: The bottleneck unit of the backbone network ResNet50 in pose-guided feature distilling generative adversarial network for robust person re-identification finally adopted a simple addition operation,which lacked the ability to select features.Aiming at this problem,this paper introduces the attention mechanism module and proposes person re-identification based on the attention mechanism.Through the learning of the channel domain and the spatial domain,the attention mask of the features is obtained to perform adaptive learning of the features,enhance the useful information and suppress the useless information.In addition,the attention mechanism and the generative adversarial network in the network can perform auxiliary training on the encoder from different angles,while eliminating redundant features such as posture,and obtaining feature representations with higher IDs.On the Market1501 dataset,top-1 reached 91.4% and mAP reached 79.1%,which was 0.9 percentage points and1.4 percentage points higher than the FD-GAN algorithm.(2)Person Re-Identification based on the aggregated residual transformations for deep neural network(ResNeXt): In order to further improve the robustness of pedestrian characteristics,we introduced the aggregated residual transformations for deep neural network to the person re-identification based on the attention mechanism above,and proposed a person re-identification based on the aggregated residual transformations for deep neural network.Compared with the deep residual network ResNet50,ResNeXt50 combines the split-transform-merge strategy of Inception and improves the cardinality(the number of transforms)without increasing the network width and depth,improved recognition results.We use this advantage to improve the model and further increase the accuracy of the model.On the Market1501 dataset,top-1 accuracy reached 91.7%.