Research On Deep Metric Embeddings And Generative Adversarial Network Based Person Re-identification Methods Under Occlusion

Person re-identification(ReID)refers to the task of retrieving a given probe pedestrian image from a large-scale gallery collected by multiple non-overlapping cameras.Person ReID has attracted spread attention in the field of computer vision and it plays an important role in many practical applications,such as intelligent security,video surveillance.With the development of deep learning,the performance of person ReID has been significantly improved.On the other hand,occlusion(such as background occlusion,pedestrian occlusion etc)is ubiquitous in natural scenes,which seriously degrades the performance of person ReID in the practical applications.Therefore,it is a critical and challenging research direction to study how to learn discriminative deep metric embeddings and how to de-occlude pedestrian images for person ReID under occlusion.The main works of this thesis are organized as follows.Firstly,most of existing person ReID methods take advantage of local features extracted from different local regions to deal with occlusions,without considering the spatial relationship among different local regions.We propose a novel person ReID method,which extracts the discriminative feature representations of the pedestrian image based on Long Short-Term Memory(LSTM),dealing with the problem of occlusions.In particular,the multi-directional spatial encoded local features are developed to learn the spatial dependencies between the local regions by taking advantage of LSTM.Meanwhile,we propose a novel loss(termed the adaptive nearest neighbor loss)based on the classification uncertainty to effectively reduce intra-class variations while enlarging inter-class differences within the neighborhood of the sample.The proposed loss allows the deep neural network to learn an embedding space and obtain discriminative metric embeddings as pedestrian representations,thus significantly improving the generalization capability of recognizing unseen person identities.Extensive comparative evaluations on challenging person ReID datasets demonstrate the significantly improved performance of the proposed method compared with several state-of-the-art methods.Secondly,in natural occlusion scenarios,the occluded image not only loses the relevant target information,but also introduces additional interference information,which makes the deep neural network difficult to learn deep robust features.Meanwhile,inspired by the powerful image generation ability of generative adversarial network on the computer vision tasks,we propose a person ReID method under occlusion based on generative adversarial network.Specifically,we firstly use the paired occluded images and original unoccluded images to training the generator and discriminator.The generator can restore the lost information under randomly occluded areas and generate high-quality reconstructed images;while the discriminator can distinguish whether the input image is a real image or a generated image.Then,we use the trained generator to generate the de-occluded training images.Adding these de-occluded images to the original training images can increase the diversity of training samples.Finally,we train a classification network based on the augmented training set.Experimental results on several challenging person ReID datasets demonstrate the effectiveness of our proposed method.