Research On Person Re-identification Using Triplets And Multi-scale Model

Person re-identification(Re-ID)is an important task in computer vision and has many applications in video-based surveillance.The application of computer vision algorithm can quickly find suspicious people in massive video,and can also be used to realize intelligent monitoring of road traffic safety.Compared with the traditional manual method of searching for specific person information in a large number of v:ideo images,the application of Person re-identification algorithm can not only save expensive labor costs,but also avoid visual fatigue,so as to find specific person information more accurately.Therefore,Parson re-identification algorithm not only has theoretical significance,but also has practical application value.Recently,the design of Person re-identification network with triplet loss has been popular in the deep learning framework.It is particularly important to note that the selection of hard triplets has significant influence on the performance of the learned deep model.However,the existing algorithms of triplet only focus on some specific forms of hard triplets,thus leading to weaker the generalization capability of model.Some researchers also use the method of segmentation of pedestrian images to study Person re-identification algorithm.The way of image segmentation can effectively learn more refined features of person,but direct segmentation often leads to the problem of inconsistency between segmentation boundaries and actual human body construction boundaries.Firstly,the current Person re-identification algorithms are summarized and analyzed,and the advantages and disadvantages of some algorithms are pointed out.Secondly,classical networks and related algorithms are introduced.Then,this paper proposes designed algorithms,and verifies the effectiveness of the algorithms on the common open source datasets Market,1501,MARS and Duke MTMC-relD.The main contributions of this paper include:Aiming at triples,this paper propose a novel variant of the triplet loss,named EHTM(exhaustive hard triplet mining loss),which is able to deal with various forms of hard triplets in a comprehensive mandistinguishing ner.Moreover,the proposed loss comprises a term to facilitate different identities by directly narrowing intra-class distances and indirectly enlarging inter-class distances.For the proposed loss,we provide an effective training strategy,named OHTS(Online Hard Triplets Selection).It further enhances model performance.A large number of basic experiments and ablation experiments show the effectiveness of the proposed algorithms.To address the problem of inconsistent segmentation,this paper proposes Multi-scale Person Re-ID Network(MPN),which can segment pedestrian images at multiple levels.To some extent,it alleviates the inconsistency between hard segmentation and actual human body structure.Furthermore,we consider the problem of over-fitting caused by label,and adopt label smoothing regularization(LSR)to alleviate this phenomenon.Extensive experiments show that designed network framework outperforms state-of-the-art approaches by a useful margin.