Design And Implementation Of Person Searching UAV System Based On Onboard GPU

In recent years,due to frequent personnel loss events caused by geological exploration,tourism exploration or geological natural disasters,time is life for lost personnel.How to complete the search of specific areas in a short time has become a hot research direction.As a flexible,portable and low-cost device,UAV is more and more applied to the search task of lost contact personnel.At the same time,it realizes independent search combined with deep learning,which greatly improves the search efficiency.However,there are still the following problems to be solved in the application of the above technology in the actual search task:image recognition mainly depends on the computing power provided by mobile devices,which are affected by CPU performance,GPU performance and heat dissipation,so that the speed of image recognition can not meet the requirements of real-time;In the case of poor light conditions such as night and fog,the recognition accuracy of target detection based on visible image is low,which can not meet the accuracy requirements of search tasks.Search route waypoints need to be manually marked by search personnel,and search waypoints cannot be automatically generated;Lack of mission management and UAV management related functions.Based on the above problems,this thesis designs and implements a lost contact personnel search system based on UAV airborne GPU.The system consists of four modules:target search,task planning,task real-time preview and background management.The target search module is responsible for searching the target in the task area.According to the specific functions,it can be divided into two sub modules:task execution control and target detection.The task execution control sub module uses the Android application developed based on msdk to send control instructions and task information to the UAV.The target detection sub module uses Darknet training samples to obtain the yolov4 tiny detection weight,and then optimizes the model through tensorrt,Finally,the powerful computing power provided by airborne GPU is used to realize real-time target detection.At the same time,by carrying a sensor system integrating multiple types of sensors,the UAV can obtain visible light images and thermal imaging images at the same time.In case of poor lighting conditions such as night and fog,the thermal imaging images can be recognized to improve the recognition accuracy in such scenes.The task planning module can automatically generate the search route covering the task area according to the task area and relevant parameters set by the search personnel.The search personnel can adjust the generated route according to the actual task scenario to meet the actual task requirements.The task real-time preview module is responsible for displaying the real-time video stream of UAV,real-time location of UAV,task execution progress and target detection results.The background management module adopts B/S architecture and uses spring boot and Vue framework for front-end and back-end development respectively.It mainly provides functions such as task management and UAV management.Relevant personnel can uniformly manage tasks and UAVs to facilitate task reuse and delegation.After the implementation of the system,the various modules and the whole system are fully tested.The test results show that the system effectively solves the problems existing in the existing UAV search system,can better assist the searchers in the search of lost contacts,and further improve the efficiency and accuracy of the search.