基于深度学习的农用无人机自主避障研究  

Research on Autonomous Obstacle Avoidance of Agricultural UAV Based on Deep Learning

作　　者：李继辉 白越[1] 裴信彪 吴和龙 Li Jihui;Bai Yue;Pei Xinbiao;Wu Helong(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,长春130033 [2]中国科学院大学,北京100049

出　　处：《农机化研究》2021年第3期1-7,共7页Journal of Agricultural Mechanization Research

基　　金：国家自然科学基金项目(61304017)。

摘　　要：针对目前农用植保无人机(UAV)自主避障能力弱及避障系统繁琐等问题,提出了一种适用于植保无人机的基于深度学习的端到端自主避障方式。利用植保无人机挂载的双目相机实时采集图像,当检测到障碍物与植保无人机距离≤5m时,自主避障系统启动,将采集图像预处理后输入卷积神经网络,输出姿态角与油门量控制无人机自主飞行与避障,同时卷积神经网络通过手动飞行采集信息进行训练。实验结果表明:该方法能使植保无人机对农田常见障碍物房屋、树木、电线杆等做出自主避障,且模型具有一定的泛化能力,适当训练后,可将此避障方式应用于复杂环境下的植保无人机自主避障。Aiming at the problems of weak autonomous obstacle avoidance ability and cumbersome obstacle avoidance system of agricultural plant protection unmanned aerial vehicle(UAV),an end-to-end autonomous obstacle avoidance method based on in-depth learning for plant protection UAV is proposed.The binocular camera mounted on the plant protection UAV is used to collect real-time images.When the distance between the obstacle and the plant protection UAV is less than 5 metre,the autonomous obstacle avoidance system is activated.After the pre-processing of the collected images,the convolutional neural network is input to control the autonomous flight and obstacle avoidance of the UAV by output attitude angle and throttle quantity.The convolutional neural network flies manually.Lines collect information for training.Experiments show that this method can make the plant protection UAV to make autonomous obstacle avoidance for common obstacles in farmland,such as houses,trees,poles,etc.And the model has certain generalization ability.After proper training,this method can be applied to plant protection UAV autonomous obstacle avoidance in complex environment.

关 键 词：植保无人机 卷积神经网络 深度学习 自主避障 

分 类 号：S252[农业科学—农业机械化工程] TP391.4[农业科学—农业工程]