Person Re-identification Based On Feature Fusion

Person Re-identification(Re-ID)is a cross-camera pedestrian retrieval problem.It is a kind of technology that determines whether there is a specific pedestrian in the surveillance record of another camera.With the actual demand for ensuring public safety and the increasing number of surveillance cameras in parks,streets and other places,it is necessary to apply Person Re-ID technology in the intelligent surveillance system.However,since the images captured by different cameras are really different in background,lighting,posture,and resolution,the task of Person re-identification is facing great challenges.In addition,the common occlusion problem in pedestrian images also increases the difficulty of pedestrian retrieval.Therefore,the keys to solve these problems of Person Re-ID are to extract robust feature representation and design appropriate discriminant methods.In recent years,deep learning methods have been widely used in Person Re-ID and achieved good results,which has greatly promoted the progress of Person Re-ID technology.A large number of existing deep learning methods are supervised learning methods,which need large-scale labeled datasets.However,the existing datasets of Person Re-ID are small in scale,which makes it difficult to simulate the real scene.Moreover,large amounts of data annotation require a lot of human resources.Therefore,in order to reduce the cost of data annotation and improve the usability and scalability of Person Re-ID in the actual scene,the research of unsupervised Person Re-ID is also a major direction of Person Re-identification research in the future.In response to the above issues,the following researches have been carried out in this paper:1)A supervised Person Re-identification method based on image-space feature fusion is proposed.Firstly,spatial features are extracted from the image features to locate the pedestrians in the image.Then,by fusing image features and spatial features,the model pays more attention to the foreground pedestrians in the image rather than the cluttered background.Secondly,the fused features are divided into several local features horizontally,and each of them is discriminated separately to solve the pedestrian occlusion problem.Next,the local features are used to train the classification loss,and the global features are used to train the triplet loss.Finally,an objective function composed of the classification loss and the triplet loss is proposed.This method can extract more discriminative features and effectively improves the performance of pedestrian retrieval.2)An unsupervised Person Re-identification method based on local feature fusion is proposed.In order to solve the problem of unsupervised metric learning,a new unsupervised triple loss is proposed.For each training sample,a hard positive is generated by random transformation such as image cropping or brightness adjustment,and a hard negative is mined by the proposed hard negative mining algorithm.The loss is not limited to network structures or the fields of research,it is a metric learning method suitable for unsupervised learning.At the same time,the loss is applied to a network structure that fuses local features.This network uses the optimization of local features to mine fine-grained information in images,which effectively improves the performance of Person Re-ID in unsupervised scenarios.The method reduces the high cost of manual data labeling,and improves the practicability of person re-identification method in the actual scene.The paper has 16 pictures,21 tables,and 78 references.