Research On Person Re-Identification For Multiple Scenes

Person re-identification is a technique that uses computer vision and image recognition methods to determine whether a particular person exists in an image or video sequence.With this technique,we can get the suspect escape route information and find lost children,which plays an important role in urban security construction.In the monitoring scene,person re-identification faces many challenges due to the influence of camera angle,lighting conditions and person pose,due to the influence of camera angle,lighting conditions and person pose.The traditional person re-identification is based on the artificially designed features,which first extracts the characteristics of person,such as color features,texture features,etc.and then measure the similarity between person features through distance metrics.However,in multi-scenarios,the artificially designed features are not robust,so the traditional methods are less than ideal.Deep learning has achieved great success in many areas of computer vision since it was proposed.The use of deep learning to solve the problem of person re-identification has also received more and more researcher's attention.In addition to the recognition of person,person attribute information recognition,such as identifying person clothes color,also plays an important role in pedestrian recognition.In multi-scenarios,person in surveillance video will face challenges such as different poses,but the attribute information of person is constant,for example,backpacks carried by pedestrians.By identifying the pedestrian attributes,the pedestrian's activity characteristics can be obtained,which is beneficial to the mall to obtain customer preference information.In response to the above problems,this thesis proposes the following methods based on the convolutional neural networks:(1)An improved loss function is proposed.The time and place of the existing pedestrian attribute data set are relatively fixed,which results in uneven positive and negative samples of the pedestrian attribute in the training set.To solve the problem of unbalanced attribute samples in person attribute recognition,this thesis is proposed to adjust the loss function weights by using the number of attributes,so that the convolutional neural network pays more attention to the attribute information with less samples,thereby improving the accuracy of attribute recognition.(2)A person re-identification network based on person attributes and poses is proposed.For different scenes with different poses,this thesis first detects the key points of the human body in the image,and then generating the person pose image by the human pose key points.The person pose image pay more attention to the person's local feature information.And it can reduce the impact of background information.Therefore,the pose image is used to extract the local features of the person,so that the network learns person feature information that may be easily ignored in the global feature.And then use person attribute information to increase person re-identification rate.What's more,a dropout layer is added to the network to prevent the network over-fitting.This method can not only re-identify person,but also recognize person attribute information.In this thesis,experiments are designed and tested on public data sets Market-1501 and DukeMTMC-reID,the proposed method has been proven to be effective through experiments.