| Recently,with the development of artificial intelligence technology,increase number of technologies are considered to be applied to assist human life in reality.Person re-identification,an important research branch in the computer vision,is expected to be used to search and match pedestrians with the same identity in the multi-camera network.This research can help human to describe the trajectory of pedestrians and retrieve specific pedestrians in surveillance video system.Therefore,person re-identification has high application value in business and security.Due to the complexity of the real environment,the great change of illumination condition and lens perspective,crowded and occlusion,person re-identification has great challenge in practical application.In order to improve the performance of it,the following work has been done in this paper:1.This paper proposes a convolutional neural network that takes into account the global and local features of human body.network combines multi-scale features to provide discriminative features for re-identification tasks.In addition,the whole network was trained with the supervision of multi-loss,which are cross-entropy loss of label smoothing and the triple loss of hard sample sampling.Experiments show that the use of label smoothing in cross-entropy loss can reduce over-fitting of model,then has a positive effect on the accuracy.The use of the triple loss of hard sample sampling improves the recognition ability of the network.In the end,the global and local features of the network are complementary to each other.The performance of the network is boost after fusion those multi-scale features.2.This paper proposes a re-ranking framework with Image Pool,which uses multiple image features of the target person as retrieval.It can make the features of the target person richer and more comprehensive,then improve the accuracy.In order to reduce the redundancy between adjacent video frames,the framework firstly takes the diversity and reliability of query image into consider,and chooses a fixed number of images from multiple images to name the target person's Image Pool.Secondly,in order to further improve the accuracy,framework computes initial ranking lists of every samples in Image Pool,and proposes a'Multiple-Image joint re-ranking algorithm'to aggregate initial ranking lists.Framework calculates the rank score of partial elements of initial ranking lists.In the end,final ranking list is obtained by ascending order of the rank score.Experiments show that this re-ranking network can further improve the accuracy. |
