Research On Person Reidentification Technology Based On Deep Feature In Surveillance Video

Person Reidentification(Person re-ID),which has practical value in pedestrian detection and tracking,is used to recognize people from different cameras.With the popularization of the surveillance video,the technology of person reidentification will be used more and will become hotspot of computer vision in the future.In recent years,deep learning has salient result in a slew of areas of computer vision,thus some researches apply it in person reidentificaiton.This paper focuses on technologies of Person re-ID that apply deep learning.There are two key parts of Person re-ID,one is feature extraction and the other is distance measure.At present,there are researches using end-to-end framework to combine these two parts,which means to output the similarity of two pedestrian images directly.But this method is not adapts to large scale online search,it requires to calculate the distance between probe image and every image in the gallery.Therefore,we only consider using deep learning to extract feature in this paper.It is allowed to apply inverted index;or some other methods to accelerate search efficiencyafter we extract image features.Classification model and Siamese model are two popular models of deep learning.The classification model is limited to the size of the Person re-ID dataset,and is difficult to do training.The Siamese model includes models that based on pair-wise input and triplet input.To some extent,this kind of model solves the problem of small dataset size,but it doesn't make full use of the ID label of the image.What's more,training samples of the Siamese model is generated offline currently,which leads to lots of inefficient pair or triplet samples.To solve problems as mentioned above,this paper have done the following work:1.Learning feature with the network which combines softmax-loss and triplet loss.The combination of two losses make full use of the information of image label and the similarity of two images,which can learn efficient feature to recognize person even under the circumstance that the training dataset is small.The lifted triplet loss be used in this paper chooses training data online,which makes full use of all image pairs of a mini-batch and gets a better result than traditional triplet loss.2.This paper designs a Person re-ID system on surveillance video.Considering datasets of Person re-ID are small,which constrains the generalization ability of the deep model,and the hand-crafted feature may not be influenced by the dataset.Therefore,this paper combines deep feature and hand-crafted feature in the system.What's more,this paper focuses on the Person re-ID problem which uses single image as input,so it applies query expansion to improve the retrieval result.This paper uses simple and useful methods to improve the result of person retrieval,which has reference value on applying the Person re-ID technology in the real world.