Researches On Person Re-Identification Based On Convolution Neural Network

In recent years,with the development of society,surveillance cameras have been widely applied in various fields.Instead of traditional manual monitoring methods which plays a huge role in maintaining social stability and fighting crimes.The problem of re-identification is the core issue of specific target video surveillance retrieval.It aims to specify the target of a specific pedestrian in different periods of video captured by different cameras.However,the existence of a series of problems such as pedestrian posture,camera Angle,shooting image resolution and so on makes the subject face this huge challenge.Usually,pedestrian weight recognition mainly includes two aspects of feature representation and metric learning.The traditional solution is to measure the similarity through metric learning after artificially designing the feature,but this way is not very good when facing the real complex scenes.Because deep learning has been widely used in various fields of computer vision,such as image classification,face recognition and other tasks,and has achieved great success,so this paper attempts to solve the problem of feature expression and distance measurement by using deep learning method.In this paper,we mainly focus on two aspects in the study of the problem of person re-identification using deep learning methods:1.On the loss function of the neural network,the loss function suitable for its use is designed according to the characteristics of the pedestrian weight recognition task.Different from the face recognition classification task can directly use the Softmax loss function training network,because the pedestrian sample image resolution is low,while similar pedestrian pose differences greatly,we proposed ranking loss function for positive and negative samples in the feature space distance are more finetuned to reduce the sample variance within the class,and it shows the recognition using traditional Softmax loss function was higher in the experimental process.2.the structure of neural network,combining with the local characteristics of the structure of STN pedestrian attitude estimation method,position in the input picture by adjusting the pedestrian transformation parameters of samples in the adaptive learning process of network training,through a number of cross entropy function optimization makes the local feature of each block is used as the full information network at the end of network training.The experiment in this paper is carried out on three open datasets.The index is cumulative sum characteristic curve to evaluate the experimental results.By comparing with other algorithms,we can find the method we propose is better than public method.