Research And Implementation Of Person Re-identification Technology Based On Deep Transfer Learning

Person Re-Identification(ReID)is a technique that uses computer vision technology to determine whether the pedestrians captured by different surveillance cameras belong to the same pedestrian.It has a good application prospect in intelligent security,criminal investigation and other fields.In recent years,deep learning approach represented by deep convolution networks have made breakthrough progress in the field of computer vision,but deep learning often requires a large amount of training data.The purpose of transfer learning is to use existing knowledge and experience to solve different but related domain problems.In view of the fact that only a small amount is used for person re-identification of labeled data in practice,on the basis of deep convolution network,this thesis studies how to use transfer learning technology to implement person re-identification.The main work of this thesis is as follows:1.Construct a small-scale person re-identification dataset,which contains 1613 images and 5647 labeled pedestrian bounding boxes for a total of 572 identities.The dataset is collected by means of a handheld camera through multiple scenes.In the process of image collection,as much as possible is included in the changes of viewpoints,illumination,resolution,occlusion and background,so as to reflect the actual application scenes and increase the diversity of scenes,it is applicable to the research on the problem of re-identification of small-scale data set in this thesis.2.Design and implement person re-identification algorithm based on deep convolution network fine-tuning.Based on the improved deep convolution network ResNet50,the transfer learning method with two-stepped fine-tuning is adopted.The one-stepped fine-tuning is performed on the pre-training ResNet50 use large-scale re-identification datasets,this ResNet50 model is obtained from the training of ImageNet.After training to get the pre-training model,the two-stepped fine-tuning is performed on the pre-training model use the small-scale dataset.Compared with two methods of training deep convolution network directly and one-stepped fine-tuning,the model achieves higher recognition accuracy,finally,the Rankl and mAP of the model on the dataset of this thesis respectively reached 83.89%,71.2%.3.Design and implement person re-identification based on deep adaptation network.This method adopted an end-to-end re-identification network based on ResNet50 as the basic structure of the deep adaptation network.The adaption transfer learning is performed based on the pre-training deep adaptation network,by freezing the convolution layer weight and fine-tuning the adaptation layer to reduce the classification loss of the target domain and the distance between the target domain and the source domain data in the feature space,so as to promote the transfer effect of the small dataset re-identification model.Compared with the method of two-stepped fine-tuning,the model achieves higher recognition accuracy,finally,the Rankl and mAP of the model on the dataset of this thesis respectively reached 85.08%,75.07%.