Study On Person Re-identification Algorithm Based On Deep Learning

Person re-identification(Person Re-ID)means that images of a pedestrian from cameras in a surveillance camera network can be automatically retrieved based on one of this pedestrian's image from another camera.Due to the differences in camera deployment and the movement of pedestrians,the appearance features of the same pedestrian can change greatly,posing a huge challenge to person re-identification.Person re-identification systems based on deep learning can effectively extract the appearance features of pedestrians,but they require a large amount of training data.And current person re-identification datasets are relatively small and cannot fully meet the need of deep training.Therefore,this thesis studied the person re-identification algorithm based on deep learning and explored the method of using generative adversarial net to extend the person re-identification datasets.The main research work of this thesis is as follows:We designed an efficient classifier on the residual learning network ResNet50 to obtain a deep person re-identification model named ResNetEx.Model code was written using PyTorch deep learning framework and experiments were conducted on the Market-1501 dataset and the CUHK01 dataset.The results of the experiments surpassed many of the latest advanced models,which verified the validity and advancement of the ResNetEx model.A feature enhancement experiment was conducted,which flipped the test image and calculated its feature,then added the flipped feature to the original test image feature.The result showed that the accuracy of the ResNetEx model on the two datasets continued to increase.The feature enhancement experiment was equivalent to extending the training data,and the accuracy improvement indicated that the training of ResNetEx model on original datasets was not sufficient,that is,the size of the datasets we used did not meet the need of deep training.Aiming at the problem that many person re-identification datasets were small and only had few pedestrian postures,the method of expanding person re-identification datasets was studied.A POSE-GAN model based on generative adversarial net was proposed,which can extend the pedestrian images of the datasets to eight standard poses.In the experiment,the generated images were used as strong label training data.The results showed that the images generated by the POSE-GAN model did not retain enough pedestrian appearance information,and using these images as strong label training data will reduce the accuracy of person re-identification.To address the problem that the generated images cannot be used as strong label training data,a LSROW algorithm was proposed to use the generated images as weak label training data,that is,assigning smoothed identity label to the generated images.Experimental results showed that the generated images as weak label training data slightly improved the accuracy of ResNetEx model on large dataset,and significantly improved the accuracy of ResNetEx model on small dataset by 7.6%.After adding weak label generated images to the original datasets,the accuracy of the ResNetEx model trained on both datasets surpassed all the models used for comparison,which fully demonstrated the feasibility of using POSE-GAN model to extend person reidentification datasets.