Design Of Reconnaissance And Surveillance System Based On Infrared Pedestrian Detection

In border control as well as in some military tasks,surveillance of key areas is done by manual operations.The existing small reconnaissance equipment is less integrated,large in size,and the targets require human interpretation,which cannot meet the practical needs.Therefore,it is very important to design a portable and automatic target detection all-weather reconnaissance and surveillance system to enhance the ability to control important target areas and effectively obtain battlefield information.The key of the IR-based portable reconnaissance surveillance system is the real-time detection of pedestrian targets from IR signals,and the deep learning target detection algorithm has outstanding performance in terms of speed and accuracy,which provides ideas for IR pedestrian detection in reconnaissance surveillance systems.The work in this thesis contains the following aspects:The infrared signal pedestrian detection method is studied,and for the infrared pedestrian detection task in the special context of border control and military operations,a method based on the improved YOLOv3 model to implement infrared pedestrian for detection is proposed,using the k-means clustering method to reconstruct the anchor box structure of the model and introducing the difference in the location of the pedestrian detection center to redesign the loss function of the model.In this thesis,two datasets of active infrared imaging and infrared thermal imaging are collected and created.Infrared thermal imaging has good recognition effect for the case of partially occluded targets,but its low imaging resolution leads to unfavorable pedestrian target detection.To solve this problem this thesis experimentally compares the effect of interpolation-based and learning-based image super-resolution algorithms on infrared thermal imaging data,and proposes the use of SRGAN methods to achieve super-resolution processing of infrared thermal maps.This thesis designs and implements a reconnaissance surveillance system that can work 24/7,and implements a hardware platform that integrates signal acquisition,transmission and processing from the hardware level,based on which the algorithm detection effect is tested.The experiments show that the m AP of the proposed algorithm for IR pedestrian detection is higher than that of Fast RCNN and other target detection algorithms,and higher than YOLOv3 algorithm by nearly 2%.The system could detect pedestrian intrusion and other important target areas in the night environment,and can meet the practical needs of reconnaissance and surveillance.