Research Of Person Re-identification Based On Mixed Loss Function

Person re-identification is to match two or more pedestrians photographed by cameras that do not have overlapping monitoring areas to determine whether they are the same pedestrian.With the popularity of the monitoring network,this technology can be applied to areas such as intelligent video and smart management in combination with target detection and tracking technologies.Person re-identification has become a hot topic,and researchers have proposed many ways to solve various problems,such as camera angle,pedestrian attitude,weather lighting changes and so on.Different from the traditional methods of feature extraction and metric learning,this thesis research the pedestrian re-identification problem through deep convolutional neural networks.In order to extract the pedestrian depth features with resolution and robustness,the following methods are studied,including:1.For the pedestrian image,due to the angle of view,posture,clothing,etc.,causing a large difference between the images of the same pedestrians.So,the training of the network is supervised by two loss functions with complementary characteristics,that is,using the hybrid based Lost convolutional neural networks to study pedestrian re-identification problems.The method can make similar pedestrian images aggregate to a center based on the correct classification of pedestrian images,thereby improving the expression ability of the features.The experiment proves that the method can effectively improve the effect of person re-identification.2.Further aiming at the problem that the accuracy of person re-identification is lower due to the similar "inter-class" image in the actual scene than the "intra-class" image,a Siamese network with bidirectional max margin ranking loss is proposed.Mainly through the joint of Softmax loss and Ranking loss supervise the network training on the training set,by the similarity value of the positive sample pair minus the negative sample pair similarity value is greater than the predetermined threshold.Constraints on "intra-class" relationships,at the same time,ensure the "inter-class" distance is greater than the "intra-class" distance,to further improve the accuracy.Experiments were carried out on the pedestrian dataset using the trained neural network model.Compared with other typical person re-identification methods,the method showed better performance and could effectively improve the recognition rate.