Person Re-identification Based On Deep Learning

In recent years,person re-identification has gradually become one of the most active research fields in computer vision,which is the focus of research in intelligent video surveillance systems.Person re-identification is a method of matching target pedestrians in a multi-camera surveillance image or video from an unsuperimposed perspective,which is essentially a pedestrian retrieval task.This thesis takes the person re-identification algorithm based on deep learning technology as the research topic,and focuses on the ability of deep convolutional neural network to extract pedestrian features in the person re-identification task,the hard sample mining problem in depth metric learning and the deep learning classification model based on person re-identification method,etc.The main research content is divided into three parts:1.Using the convolutional neural network to extract pedestrian characteristicsThe performance of the person re-identification algorithm depends on a robust and recognizable pedestrian feature.Compared with the traditional pedestrian feature extraction method,the deep learning model can adaptively obtain features with strong expressive ability through the gradient descent algorithm,thus avoiding the complex feature design process.In this paper,the AlexNet model and the ResNet model in the convolutional neural network are used as the basic backbone network.The dataset training optimization network is used to obtain the model with strong pedestrian feature extraction ability.The experiment proves that the feature extraction of the deep network model is better than the artificial feature design method.2.Hard sample mining based on depth metric learningThe person re-identification method of depth metric learning can only use the label information between samples,and there are often problems in the model training process that are difficult to sample.Hard sample mining is a key factor in improving the performance of the algorithm.This paper analyze the existing sampling methods,generalize the point-to-point sampling method to the point-to-point sampling method,and further assign different weights to each sample point in the point set according to the Euclidean distance between the samples.Compared with the hard sampling method,this method can adaptively judge the difficulty of each sample.In addition,this paper further improves the difficult sample weight assignment algorithm in the hard perceptual loss function,so that the loss can make the model obtain better optimization results.3.Improvement of person re-identification classification modelThe deep classification model can use the label information of the sample to supervise the training network model learning,which helps the model to extract the discriminative features.In order to obtain better recognition effect,the Softmax loss introduced by the interval mechanism is applied to person re-identification.The algorithm also finds the interval between different types of pedestrians while looking for the pedestrian feature interface.In addition,we supervise the training network by combining metric model loss and classification loss.This multi-loss joint learning approach compensates for the inadequacy of the classification model.