Research On Pedestrian Re-identification Technology Based On Deep Learning

With the rapid development of smart cities and public security,the application of video surveillance systems is being expanded.Facing the large amount of data generated in surveillance video every day,the traditional video analysis technology has been unable to meet the need of people.How to efficiently obtain the goals of interest from the video data has gradually become the focus of the research.At the same time,due to the outstanding performance of deep learning frameworks in target detection and recognition tasks,many scholars are trying to apply its frameworks to the research of intelligent video surveillance technology,and train more generalized models to solve the actual monitoring problem.Based on the above background,we devote to detect and re-identify multi-motion targets in multicamera using deep learning method in the paper.The main work follows as below:Through the research and comparison of several general target detection frameworks based on deep learning,finally,we choose the Faster R-CNN target detection framework as the research frame in this paper.We find that it is not good in speed and accuracy of small targets detection,when the framework is directly applied in the actual monitoring scene.After the visual analysis of the shallow and deep features of the convolution neural network,the concept of feature cascade is introduced to adjust and optimize the network.At the same time,in order to improve the accuracy of the trained network classifier,the difficult case mining strategy is selected to optimize the network training.Finally,the training and testing is carried out on our dataset and public dataset,and precision and speed of detection are improved.The experimental results show that the model is robust and lays a good foundation for pedestrian re-identification.At the same time,the input of the previous re-identification network is the pedestrian sequence that has been detected by the detection network.Starting from the original video sequence data,this paper integrates the pedestrian detection network and the re-identification network into a unified network.Based on the above-mentioned improved Faster R-CNN detection network framework,the similarity measurement correlation algorithm is introduced to design the end-to-end network that combines detection and re-identification functions.In this paper,we compare the effect of several similarity measurement functions in the network,improves the loss function of the basic network,and puts forward a training method for the improved end-to-end network.In terms of the re-recognition evaluation standard,the performance of the network is improved.Finally,the pedestrian re-identification system is designed by using Matlab GUI module.When the target ID is input,through the improved detection network,the targets information is displayed,and then using the re-recognition network,the top 10 targets are output.With the help of the evaluation criteria.final evaluation results will be decided.