Person Re-identification Method Based On Deep Learning And Unsupervised Learning

The development of new smart cities is an important national strategy,which is based on constructing a new generation of intelligent infrastructure with a “city brain” as the core,to create a safe and convenient smart life.The analysis and efficient use of massive cameras and surveillance video data in the city is one of the core keys to realizing the new smart city.This article mainly studies person re-identification,which re-identifies the most critical pedestrians in the video collected by a multi-camera network with a non-overlapping field of view.This technology can search and match a given target person image for person analysis,retrieval,and tracking across cameras.It is also the core technology to realize the smart city and intelligent security and has become a research hotspot in the field of computer vision.However,there are some problems with the pedestrians in the video such as occlusion,low resolution,and large differences in posture and angle of view in complex actual scenes.It is difficult to effectively extract the distinguishable features.Moreover,in the new scenario,it takes a lot of energy to obtain annotated data,and it is difficult to realize a large-scale application.Therefore,this paper focuses on the feature learning of person images in person re-identification problems and person re-identification based on unsupervised clustering methods to learn the distinguishing features of person images and solve the problem of data annotation in practical applications,aiming to establish an efficient and feasible method of person re-identification.The main research results are as follows:(1)A multi-granularity feature fusion person re-identification method based on attention mechanism is proposed.It is used to solve the problem of learning global and local features of person images in the person re-identification.First,the attention mechanism is applied in the deep learning network by using the convolutional attention module,which could simultaneously learn channel attention and spatial attention to adaptively extract pedestrian image features;Then,apply a multi-branch network to extract global features,mediumgrained features,and fine-grained features for person images simultaneously;Finally,its multi-granularity features are fused to make better use of global and local information to obtain the final robust and adaptable person image features.Great performance improvements have been made on the Market-1501 and Duke MTMC-re ID databases.It shows the effectiveness of attention mechanism and multi-granularity feature fusion in image feature learning.(2)A person re-identification method based on multi-scale feature fusion and multi-task learning is offered to use the multi-scale information of person images to improve the accuracy of person re-identification.First,using high-resolution network HRNet to extract features from person images with a deep network structure of parallel multi-branch crossover,which can learn low-resolution features of images and retain high-resolution features;Then,exchange information between feature branches of different resolutions,and use the multiscale information of the image through feature fusion;Finally,in network optimization,classification tasks and metric learning tasks are both used to perform multi-task joint learning optimization to better learn feature space mapping.Experiments on two public datasets,Market-1501 and Duke MTMC-re ID,prove the good performance of using the HRNet network for multi-scale feature fusion and multi-task learning on person re-identification.(3)A method of unsupervised hierarchical clustering person re-identification based on information entropy guidance is proposed,which is used to solve the difficult problem of labeling pedestrians in person re-identification.First,for bottom-up hierarchical clustering,the Renyi information entropy distance criterion is used to measure the distance between two clusters.When merging clusters,select the clusters with the smallest Renyi information entropy distance to fully utilize the information of all data in each cluster;Then,the center loss is introduced to measure the distance between elements in the class and the cluster center to make the same kind of data more aggregated.Finally,iteratively perform the hierarchical clustering process and feature extraction process to optimize.Experiments on Market-1501 and Duke MTMC-re ID datasets show that the use of distance criteria based on Renyi information entropy and the use of center loss have significantly improved the unsupervised person re-identification method.