Person Re-identification Based On Metric Learning

Person re-identification is to detect whether or not there is a specific pedestrian sample in photos or videos from cameras in large-scale public places from different angles.The current pedestrian re-identification technology is supposed as a branch of the image retrieval.It is widely used in areas such as intelligent video surveillance,criminal investigation,smart security and human-computer interaction.Cameras are different in their hardware and also different in the surroundings.Besides,the appearance of the pedestrian is easy to be affected by the pose,angle,shelter and the change of the illumination.All these factors make the pedestrian reidentification a research area which is worth deep digging,and pedestrian re-identification is also one of the most challenging and popular task currently.There are two procedures in the usual approach for pedestrian reidentification.First of all,a robust pedestrian feature descriptor is designed which can reflect the differences between various pedestrians for a long time.Then,using the metric learning to measure the similarity,which distinguishes the samples of the pedestrian.The traditional approach for pedestrian re-identification is to adopt the single manual feature which is not robust,so that it can't reflect the feature difference well.In the complex conditions for pedestrian re-identification,the single manual feature can't have a good effect.In recent years,multiple manual feature combinations and deep features have been widely used in computer vision areas such as the pedestrian detection,the face recognition and the image retrieval.This paper provides three methods of metric learning based on the robustness of the pedestrian re-identification feature and the difference between pedestrians,which is aimed to increase the robustness of the feature descriptor and the difference of the pedestrian identification.The research contents of this paper include: 1.Proposing a pedestrian re-identification method based on matric learning and the feature fusion which is based on pedestrian sample block.It is to use the image enhancement algorithm to adaptively enhance pedestrian samples,and then to block the pedestrian samples and extract manual features by biological proportion and to perform the feature fusion.Then it is to use PCA to reduce the feature dimensionality and finally,to use KISSME algorithm in metric learning to perform the pedestrian re-identification.The KISSME algorithm is more accurate than the algorithm for single manual feature.2.Proposing a method of pedestrian re-identification which is based on metric learning and multi-feature fusion,which can reflect the inner feature of the whole pedestrian better and increase the difference between the feature of the pedestrian.This method is to fuse the WHOS feature which is based on the rough stripe pooling with the block fusion feature above,and to use KISSME metric learning algorithm to perform the pedestrian reidentification,this algorithm is more accurate than the algorithm for multifeature fusion.3.It is believed that using the neural network which is trained by massive data to extract features is apparently more robust than using the manual features.To make the neural network trained by the non-pedestrian database fit the research for the pedestrian re-identification well,this paper provides method which is based on migration learning and deep feature learning.Under the keras deep learning framework,this method first applies the weight migration and fully connected layer replacement to the pretrained VGG16 neural network on the ImageNet dataset.Then it tweaks the the network with the augmented CUHK03 pedestrian database and with the migration learning ideas.It adopts the penultimate fully connected layer of VGG16 neural network as the feature layer,and uses XQDA metric learning algorithm to perform the pedestrian re-identification.This algorithm is much more accurate than the previous algorithm framework.The pedestrian re-identification algorithm in this paper is tested with the two public pedestrian database VIPeR and Market-1501.It uses Rank value as an index to verify the results of the test.Compared with other typical pedestrian re-identification algorithm,the algorithm in this paper performs better in recognition accuracy than most existing algorithms.