| Pedestrian re-identification(Re-ID)is the detection and recognition of target pedestrian images across cameras,and is one of the most difficult problems of computer vision.In recent years,most pedestrian Re-ID methods are based on the idea of deep learning,which uses convolutional neural networks(CNN)to extract pedestrian features from images,further improving recognition accuracy and efficiency.However,deep learning-based methods often require a large amount of high-quality labeled data for model training,otherwise over-fitting problem is likely to occur.In order to solve this problem,this paper proposes two methods from different angles.Aiming at the problem of over-fitting due to the lack of annotated data about lighting changes and posture changes in the dataset,this paper proposes a pedestrian re-identification method based on style transfer and pose transfer.First,the Cam Style model is used to generate multiple corresponding images of different camera styles to achieve data augmentation based on style transfer.Then,the pedestrian’s conditional posture and target posture are extracted according to the human posture estimator,and the Pose-Attention Transfer Network(PATN)is used to implement data augmentation based on pose transfer.Finally,new synthetic dataset after style transfer and pose transfer is used as the input of the pedestrian re-identification model for training.Experiments on the Market-1501 and Duke MTMC-re ID datasets show that this method can be better generalized for problems such as posture changes and lighting changes.So this method is more suitable for situations where the foreground image changes little and the pedestrian pose changes much.In order to be suitable for the video environment with many changes in the foreground of the characters and the appearance of occlusion,this paper proposes a Multi-stage Spatial and Appearance fusion style transfer(Camstyle)pedestrian re-identification method based on the Generation of Adversarial Networks(MSAC-GAN).In each stage,it consists of the addition of foreground images,spatial synthesis and appearance synthesis.First,add the foreground image;then use the Spatial Transformer Model(STM)to spatially adjust and scale the foreground image to obtain the spatial synthetic image;next use the appearance synthesizer to adjust the foreground image color in the spatial synthetic image to obtain space and appearance composite image.In the next stage,the output image of the previous stage is used as the background input of this stage,and an other foreground image is added.By analogy,more diverse synthetic images are obtained through multi-stage space and appearance fusion conversion.Finally,when all the foreground images are added,the dataset of the last synthetic images is input into the Cam Style model,and the amount of data is increased again.Experiments show that this method can improve the robustness of pedestrian reidentification model to occlusion. |
