Occluded Person Re-identification Method Based On Multi-scale Features And Unsupervised Data Enhancement

Person re-identification(Re-ID)technology can search specific pedestrians,which is widely used in the current intelligent video surveillance system to assist people in catching criminals or looking for missing persons.Occluded Re-ID task is proposed mainly because people are often occluded by various obstacles in the real world,which greatly affects the accuracy of model matching.At present,this task still faces great challenges.In the early stage,most of methods only focused on matching with human body without considering occlusion.Later,Some new methods based on part features appeared,which improved the generalization performance of the model,but most of these methods need to rely heavily on the strict alignment of human.When there is severe occlusion,the performance is still poor.What’s more,partial Re-ID methods need manual cutting when facing occlusion,which is very timeconsuming and laborious.Moreover,the current datasets for occlusion problems are often small-scale,which is not only easy to cause over fitting during training,but also the image style differences between different datasets make it very difficult to use these data sets jointly.It can be seen that the research about occluded person re-identification is not perfect at present,and overcoming these difficulties is of great significance for this research.Based on this,this thesis studies Re-ID technology under occlusion.The main work and contributions are as follows:(1)Aiming at the incompleteness of human body under occlusion,an occluded person ReID method based on multi-scale features is proposed.Different from manual clipping in partial Re-ID method,a partial human body locator is constructed by using the target detection algorithm to automatically recognize and cut part of the human body in this method.Then this method designs a horizontal pyramid pooling strategy to extract multi-scale features and enhance the robustness of the model under the occlusion problem.Comparative experiments show that,this method has better matching accuracy in the occluded Re-ID task.(2)In order to solve the problem that it is difficult to align the local features between different pedestrian images under occlusion,an occluded Re-ID method based on human feature reconstruction is proposed.Different from the current mainstream methods which rely heavily on local feature alignment,this method is an alignment-free approach.Based on Method 1,the human feature reconstruction distance is proposed to improve triplet loss with batch hard mining of the model.In this method,the sparse representation method is used to reconstruct human features to increase the proportion of similar parts of matching correlation.Experiments show that this method can effectively improve the occlusion resistance of the model.(3)Considering that the small-scale of labeled data for the occluded Re-ID problem,an occluded Re-ID method based on unsupervised data enhancement is proposed.On the basis of Method 2,this method adopts unsupervised method for data enhancement,adds unlabeled data and generates pseudo tags.A mixed memory model and a clustering reliability evaluation standard are proposed to update the category prototypes needed for different types of data dynamically,and the parameters of Method 2 are fine-tuning by using the unified comparison loss function.This method can not only expand the amount of data,but also reduce the over fitting of the model.Experiments show that this method further improves the performance of the model under occlusion.