| With the development of intelligent monitoring,a large amount of monitoring data is generated.By manual inspection,the efficiency is very low and the accuracy is difficult to guarantee.The analysis of monitoring data by using the person re-identification method can overcome these shortcomings,therefore,it has received more and more attention in recent years.However,most person re-identification methods focus on visible light images,visible images lack information in poor lighting conditions,and it is difficult to monitor all-weather surveillance scenes.In addition,most studies focus on the person re-identification algorithm itself,less research on pedestrian detection in person re-identification systems.In this paper,the infrared image and the visible light image are combined to study the related theories and methods of the all-weather person re-identification system.The main work is as follows.1.A multi-spectral pedestrian detection method combining infrared image and visible light image is studied.Based on the YOLOv3 target detection framework,a medium-integrated network structure is designed to fuse infrared images and visible images.The features of infrared and visible images are extracted by two different convolutional neural network branches,in order to improve the pedestrian detection effect of the all-weather scene,imaging of infrared images in night scenes is better,visible images are better imaged in daytime scenes,and the illumination information sensing sub-network is introduced to learn the features extracted by the bifurcation network,and the weight of the illumination information is obtained.The classification and regression scores are weighted by illumination information weights on three scales.Finally,the loss rate of 29.13% and the running time of 0.21 s are obtained by experiment on the KAIST dataset.Compared with other algorithms,the operation time of this method is reduced by 0.27 s with slightly higher precision,The detection effect and operating efficiency are better.2.A cross-modal person re-identification method for infrared pedestrian images and visible pedestrian images is studied.Aiming at the problem that the two different modes of infrared image and visible image are different in feature,the Res Net-50 residual network is used as the backbone network,and a two-way network is designed to extract the features of infrared image and visible image and share it.The convolutional layer parameters learn different modal information of infrared images and visible images;in order to solve the problem that the distance of the same pedestrian in different modes is greater than the distance of different pedestrians in the same mode,on the basis of the person re-identification reordering loss,The cross-modal sorting loss and the sorting loss in the modal are introduced.In order to further utilize the pedestrian's identity information,the pedestrian's identity loss is used to constrain.Finally,the experimental results show that the Top20 hit probability of the SYSU-MM01 and Reg DB datasets reached 81.8% and 74.81%,respectively,and the m AP reached 26.49% and 35.32%,respectively,and achieved good recognition results.3.Designed and implemented a person re-identification system that monitors all-weather intelligence.In order to verify the application effect of the above algorithm in all-weather scenarios,a person re-identification system consisting of a pedestrian detection module and a person re-identification module was designed and implemented.The designed system is based on the Py Torch machine learning library and the Open CV computer vision algorithm library,and is programmed in C++/Python language.After reading the camera video or the saved monitoring image,the system passes through the multi-spectral pedestrian detection module and cross-modal pedestrians.The identification module processes and obtains the final person re-identification result,and the implemented system completes the function of the all-weather monitoring system. |
PDF Full Text Request