Research On Person Re-identification Method Based On Foreground Segmentation And Multi-loss Ensemble

Nowadays,with the improvement of the computer's supercomputing power and huge storage capacity,as well as the society's need for security,monitoring devices are installed in different scenarios in various places.At the same time,the development of artificial intelligence has made people unsatisfied with the method of manually filtering videos,and the intelligent search video method is particularly important.Person Re-identification is a research hotspot of computer vision and an important technology in intelligent video surveillance.Due to the incorrect detection frame,the obstruction of obstacles,the difference in pedestrian poses and the difference in camera angles,phenomena such as misalignment of the body of the same pedestrian,different backgrounds of the same pedestrian,and partial body occlusion of the pedestrian,etc.,which brings great challenges to person re-identification.Therefore,on the basis of citing the deep network model,this paper improves the network structure and loss function,and proposes a person re-identification method based on foreground segmentation and multi-loss ensemble.The main research contents of this article include:1.Arrange and summarize the common methods of pedestrian re-identification.From the perspective of metric learning,the metric loss functions commonly used in person re-identification are elaborated in detail,including Contrastive loss,Triplet loss,Quadruplet loss,and Triplet hard loss with batch hard mining(Tri Hard loss),etc.From the perspective of local learning,enumerate the methods of obtaining local features,including horizontal stripe cutting,grid cutting,and so on.At the same time,this paper systematically analyzes various methods in metric learning and local learning.2.Due to obstructions,different camera angles and other reasons,the same pedestrian background is very different,and pedestrian parts in the image are missing.To address these issues,a network based on foreground segmentation and local alignment is proposed.Foreground segmentation is performed on the images in the dataset to enhance the influence of foreground targets on features and weaken the interference of background noise.At the same time,for pedestrians whose bodies are misaligned butthe same pixel position corresponds to different body parts,the foreground images are locally aligned.To consider accurate segmentation,the joint point positioning method is used to effectively compare the similarity of the divided components.In addition,for the problem of partial body occlusion,this paper uses a weight adaptive method to flexibly assign local loss weights to obtain a higher loss description model.3.In view of the fact that similar pedestrians in the pedestrian classification task are prone to misclassification,the multi-loss ensemble network is proposed to identify and classify pedestrians by integrating local loss,global loss and background?bias constraint loss.The network model is mainly composed of three parts: foreground segmentation and local alignment net,global-loss ensemble net and background?bias constraint net.First,the foreground segmentation and local alignment method is used to segment the foreground and adaptively weight the local loss to solve the problem of missing body parts.At the same time,global losses such as foreground loss and panoramic loss are used to enhance the generalization ability of the model.In addition,the background?bias constraint net is to reduce the influence of background information again by limiting the feature distance between foreground,background and panorama.Finally,it integrates multiple losses such as global loss,local loss and background?bias constraint loss.In this paper,the network is improved and optimized on the basis of the triple measurement model.This article evaluates the method on the Market-1501,Duke MTMC-re ID and CUHK03 public datasets.A large number of experimental results show that it can reduce the probability of misclassification of similar categories and improve the recognition rate.