On Pedestrian Re-identification Algorithm Based On Multi-Classification Constrained Covolution Neural Network

Pedestrian re-identification is an important subject of intelligent monitoring and computer vision,and it is widely used in the field of video surveillance.It is a technology that uses computer vision technology to judge whether there are specific pedestrians in an image or video sequence.It is generally used in the monitoring scene of non-overlapping cameras.With the development of the research,person re-identification has great practical value in solving cases,unmanned driving,pedestrian trajectory analysis and so on.After a lot of investigation and experiment,some improvements have been made to some of the imperfections in the method of pedestrian re-identification,and a convolution neural network based on multi-classification constraints has been designed for the existing problems,which is used in pedestrian re-identification tasks.The main content includes four parts.One is to optimize the convolutional neural network for the existing problems;the other is to specify the optimized algorithm implementation process;The third is to combine the two directions of pedestrian detection and pedestrian recognition.Fourthly,the effectiveness of the algorithm and the application value in the actual scene are proved by experiments.Most of the methods based on deep learning can get higher recognition accuracy,but many models are designed to compare the similarity between images at the end of the network.The difference between images at other levels is not considered,and more prominent local modules can not be utilized.In this thesis,through visualization of low-level feature map,it is found that there are significant differences between positive and negative sample pairs.Therefore,a deep neural network structure with multi-classification constraints is designed.In order to enhance the local similarity,two feature difference layers are added between the lower layer and the higher layer of the twin network,and the loss function is used to optimize the classification task.In order to learn more discriminative pedestrian features,random noise and random clipping are added to enhance training data and increase the diversity of samples.Multi-level joint training is carried out on the extracted low-level and high-level features.And the performance of the model is improved by changing the training strategy and replacing the loss function after the specified number of iterations.The test results of two large data sets CUHK03 and Market-1501 indicate that the network structure designed in this thesis is better than most current methods.Since most researches focus on improving the accuracy of re-identification and neglecting real-time performance,in this thesis a pedestrian detection re-identification system is designed by combine YOLOv3 and pedestrian re-identification network,which is capable of real-time detection.The target tracking algorithm is used to solve the problem that the pedestrian detection efficiency is low,the real-time performance cannot be satisfied,and the occlusion is present.Finally,the occlusion problem is solved,and at the same time,the model test frame rate is doubled.Then the pedestrian recognition re-identification network is applied in the actual monitoring scene.Experiments show that the pedestrian detection and re-identification system can basically meet the requirements of real-time on the basis of ensuring accuracy.