Research On Person Re-identification Based On Multi-scale Feature Combination

With the increasing demand for a better life of the people,Person Re-identification(Person Re ID)technology has become a powerful security guard,which is often used in the field of intelligent monitoring and smart life.Person Re ID technology is a hot research topic in the field of computer vision,which means to retrieve specific person images across cameras and scenes.Currently,due to the influence of camera parameters,shooting conditions,shooting time,person attitude change and other factors,it is easy to cause the poor quality of the captured person images,which makes the Person Re ID technology faces enormous challenges.Therefore,there is a long way to go to research a method with high recognition accuracy.Relying on deep learning technology,this paper improves the network framework of Person Re ID from two aspects: feature extraction and metric learning.The main research contents are as follows:(1)In order to solve the problems of network model complexity and low identification rate,a Person Re ID method based on global feature stitching is proposed.Firstly,it extracts the global features using convolutional neural networks(CNNs).Secondly,it stitches the different spatial scale features from different convolution layers together to complement the feature information.Finally,it convolutes again to obtain the features with high representation ability.In network training stage,it combines the cluster loss with label smoothing loss,and adopts random erasing augmentation(REA)as well as pooling step reduction techniques.(2)In order to solve the problems of insufficient extraction of effective feature information and weak model generalization in the first method,and the hope that the network model framework will be simpler.On the basis of the first method,a Person re-identification method based on multi-branch feature fusion is proposed.Firstly,each of the last 3 convolution blocks is connected to a respective branch.Secondly,use the batch feature erasing(BFE)model to extract local features for one of the branches,and approaches such as attention mechanism,regularization are used to deal with the other two branches to extract global features.Finally,the feature of each branch is fused to obtain the high fine-grainted representational feature.The 3 branchs monitor each other during training.A joint multi-dataset training strategy is adopted to improve generalization ability.In the end,many verification experiments were carried out on the datasets of Market1501,Duke MTMC-re ID,CUHK03,MSMT17 and Rand Person.Experimental results show that the first method can significantly improve the recognition accuracy using global feature splicing,joint loss function training and so on,for instance,the Rank-1,m AP and m INP on CUHK03 are 69.1%,66.9% and 46.2%,respectively.The second method that fuses multi-scale features and joints multi-dataset training strategies can once again improve the recognition accuracy and strengthen the generalization ability,for instance,the Rank-1,m AP and m INP on CUHK03 are 77.4%,75.1% and 51.5%,respectively.