Research On Occluded Pedestrian Re-identification Technology Based On Deep Learning

Pedestrian Re-Identification(Re-ID)is a sub-problem in the field of image retrieval that employs computer vision techniques to assess whether a specific pedestrian appears in an image or a video sequence.Deep learning has advanced fast in computer vision tasks in recent years,and deep learning-based pedestrian Re-ID approaches have been widely used in intelligent security,criminal identification,intelligent transportation,and other areas.However,pedestrians occluding each other and occluded by obstructions,inevitably limit the application of pedestrian Re-ID in actual scenes and pose serious challenges.This thesis proposes some solutions to the problem of occluded pedestrian Re-ID in the occluded environments,including the following three points:First,to address the problem of the insufficient number of pedestrian Re-ID datasets and pedestrians occluded by obstructions that decrease the re-recognition accuracy,this thesis proposes a Pedestrian Image Inpainting Generative Adversarial Network(PIIGAN),which considering the availability of pixel information of image itself.The algorithm consists of a generator network and a discriminator network.The generator network uses a codec structure to capture the overall structural-semantic information of an image through a jump connection layer and a translation layer.It translates the embedded feature of known regions to unknown regions for inpainting and uses guidance loss to minimize the distance between the decoder features and the encoder features of the obscure images.The experimental results show that the pedestrian images inpainted by PIIGAN have clearer semantic and finer texture features than some traditional image inpainting algorithms.The inpainted images can be used to expand the dataset and also improve the accuracy of the re-recognition task.Second,to address the shortcomings of the triplet loss function in metric learning,this thesis proposes a centroid loss function,which extends its point-to-point computation into a pointto-set computation and suppresses possible outliers by aggregating sample sets.Then to address the problem of difficulty in extracting discriminative features from occluded pedestrian images,the thesis proposes a Joint Non-local Attention and Centroid Loss Network(JACL-Net).To learn the relationship between pixels at different distances,the non-local attention network is integrated into the backbone network.Then the method combines the centroid loss with the triplet loss to learn discriminative pedestrian features.Finally,the network extracts global features and matches features by using the similarity metric.The experimental results show that the proposed JACL-Net method obtain the Rank-1/mAP metrics exceeding the benchmark model by 13.50%/10.30% on the occluded dataset Occluded-Duke and by 8.20%/10.2% on the occluded dataset Occluded-ReID.The Rank-1and mAP metrics of the JACL-Net method are also improved compared with the performance of existing state-of-the-art methods on the occluded and standard datasets.Experimental results show that the algorithm can effectively extract high discriminative features of images on occluded or standard datasets.It also has superiority over other existing state-of-the-art methods.Third,based on the above research of occluded image inpainting and feature extraction algorithms,to further verify the performance of the proposed Re-ID algorithm,a pedestrian Re-ID intelligent monitoring system is designed and implemented in this thesis.The system retrieves a pedestrian image uploading by the user,extracts the features by the local model of pedestrian Re-ID,matches the image with the pedestrian features in the database by using the similarity metric,and finally retrieves the target person.Experiments verify that the proposed Re-ID method in this thesis has good performance and the system has high practical value in real life.