Person Re-identification Based On Convolution Neural Networks And Gaussian Mixture Model

In recent years,as an important means of security field,the demand for security in intelligent video surveillance system is increasing with each passing day.Person re-identification is an important technology in intelligent surveillance system.At present,it has become a hot topic in the field of image processing.Person re-identification refers to the technique of locating and identifying pedestrians under one camera,and then searching and identifying the pedestrians under the cameras in other different areas.It is still a challenging problem to improve the recognition rate of pedestrians recognition technology because of various factors such as illumination change,occlusion,shooting angle and so on.At present,deep learning has been successfully applied in many fields,such as image detection,target recognition,image classification,speech recognition and so on.It also has a certain degree of research in the field of person re-identification.Compared with the traditional feature extraction method,the deep learning network can automatically learn more robust and representative features from the images.Therefore,based on the theory of deep learning and neural network,a person re-identification method based on convolution neural network and gaussian mixture model is proposed in this paper.In 2017,in the CVPR(Computer Vision and Pattern Recognition),a IDE person re-identification algorithm was proposed by Liang Zheng et al.This method is an improvement on IDE algorithm.Firstly,all the pedestrian images in the dataset are sent to the CNN(Convolutional Neural Network)for feature extraction.Then,the K mean clustering algorithm is used to cluster the extracted features to generate dictionaries.According to the vocabulary in the dictionary,the GMM(Gaussian Mixture Model)is constructed,in which each word in the dictionary is the mean of each Gauss,and the weight of each Gauss in the mixed Gauss is quantized to the sample number of each word to account for the proportion of the total sample.Then the extracted CNN features are replaced by the gaussian mixture model for feature representation.The specific method is to obtain the probability density of the feature at each Gauss,and then use the probability density and the corresponding weight of Gauss in the gaussian mixture model to get a set of probability values,and the probability values of the group are arranged in descending order.The Gauss with the largest index number is obtained,and the CNN feature is expressed by the mean of this gauss,and the quantitative characteristics are obtained.The video sequence characteristics of all pedestrians under each camera are extracted,and the characteristics are the mean value of all the images in the corresponding video sequence.Finally,the video sequence characteristics of pedestrians under the two cameras are taken as input,and the XQDA distance metric learning is used to determine whether the corresponding pedestrians are the same pedestrians.The experimental results show that the person re-identification method based on the convolution neural network and the gaussian mixture model algorithm has a higher pedestrian weight recognition rate compared with the IDE algorithm.In the same two cameras,the CNN feature extracted from the pedestrian video sequence has been processed by the gaussian mixture model proposed in this paper,and the recognition rate of pedestrian quantization features is higher.The recognition rate of person re-identification is nearly 8%?9%higher than that of directly extracting CNN features and recognition,and the highest recognition rate even reaches 86.76%.Therefore,the person re-identification method based on the convolution neural network and the gaussian mixture model can be well applied to the person re-identification technology in the monitoring video system.