基于PNP的多尺度目标视觉智能辅助降落系统  

PNP-based Multi-scale Target Vision Intelligent-assisted Landing System

作　　者：夏路 王琳[1] 梁颖茜 Xia Lu;Wang Lin;Liang Yingxi(AVIC Xi’an Flight Automatic Control Research Institute,Xi’an 710076,China)

机构地区：[1]航空工业西安飞行自动控制研究所,陕西西安710076

出　　处：《航空科学技术》2020年第10期21-29,共9页Aeronautical Science & Technology

摘　　要：小型无人机系统一般有气压计、陀螺仪、加速度计、光流计和GPS等机载传感器,由于飞行过程中机体结构产生的振动、建筑物遮挡和电磁干扰等因素,无人机的姿态和位置信息测量包含严重的测量噪声,无法满足精确降落要求,为保证无人机自主降落精度、实时性和安全性,利用计算机视觉技术和深度学习框架设计多尺度目标视觉智能辅助降落系统,针对GPS定位系统受到严重干扰情况,通过多尺度合作目标和PNP算法进行精确姿态和位置解算,从而向无人机系统提供精确的降落导航信息,在50Hz仿真条件下,视觉智能辅助降落系统在距目标20cm处能提供0.7cm/1.2°的定位精度,距目标15m处能够提供12cm/3°的定位精度,平均精度约为3cm/2°,能够满足无人机自主降落要求。Small drone systems generally have airborne sensors such as barometers, gyroscopes, accelerometers,optical flow meters, and GPS. Due to the vibration of the airframe structure, the blockage of buildings, and electromagnetic interference, the measurement information of drone attitude and position contains serious noise,which cannot meet the accurate landing requirements. In order to ensure that the drone has sufficient accuracy, realtime and safety when autonomously landing, for the situation where the GPS positioning system is severely disturbed,this paper uses computer vision technology and deep learning framework to design a multi-scale target visionassisted landing system, using the PNP algorithm and multi-scale cooperative targets for accurate attitude and position calculation, and supply accurate landing navigation information to the drone system, and under the simulation condition of 50 Hz, the vision-assisted landing system can provide 0.7 cm/1.2° positioning accuracy at 20 cm distance from the target,15 m away from the target can provide 12 cm/3 ° positioning accuracy, the average accuracy is about3 cm/2°, which can meet the autonomous landing requirements of drones.

关 键 词：计算机视觉 深度学习 合作目标 自主降落 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]