基于Ghost改进的YOLOv5轻量化双目视觉无人机避障算法  

Improved YOLOv5 lightweight binocular vision UAV obstacle avoidance algorithm based on Ghost module

作　　者：贾一凡 曹天一 白越[1] JIA Yifan;CAO Tianyi;BAI Yue(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China;SWJTU-Leeds Joint School,Chengdu 610097,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,吉林长春130033 [2]中国科学院大学,北京100049 [3]西南交通大学-利兹学院,四川成都610097

出　　处：《液晶与显示》2024年第1期111-119,共9页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.11372309,No.61304017);吉林省科技发展计划重点项目(No.20150204074GX,No.20160204010NY);省院合作科技专项资金(No.2020SYHZ0031);中国科学院青促会项目(No.2014192);中国科学院轻型动力创新院重点基金(No.CXYJJ20-ZD-03)。

摘　　要：为解决无人机在室外实际飞行时的自主避障问题,提出一种基于Ghost改进的YOLOv5轻量化双目视觉无人机避障算法。首先,引入Ghost模块改进YOLOv5中的CBL和CSP_X单元,使用CIOUloss作为回归损失函数,并将非极大值抑制CIOUnms修改为DIOUnms以优化损失函数;其次,对双目相机进行标定和校正;使用ORB特征点提取和滑动窗口匹配算法得到检测目标的视差值,再根据视差值和相机内参求解出障碍物的距离信息;最后,根据障碍物的位置和距离实现无人机的自主避障。该避障算法在嵌入式系统中运行的平均FPS达到14.3,并用无人机避障飞行试验证实了该算法的可行性;改进后的网络检测平均准确率为76.88%,与YOLOv5相比,平均检测精度均值下降0.37%,但检测时间下降22%,参数量下降25%。该算法对无人机的自主避障具有重要的应用价值。To address the issue of autonomous issue of autonomous obstacle avoidance during unmanned flight in outdoor environments,a lightweight binocular vision-based UAV obstacle avoidance algorithm was proposed utilizing Ghost module to improve YOLOv5.Firstly,the Ghost module was introduced to enhance the CBL and CSP_X units of YOLOv5,while utilizing CIOUloss as the regression loss function,and optimizing the loss function by modifying the non-maximum suppression from CIOUnms to DIOUnms.Secondly,the stereo cameras were calibrated and corrected,and the ORB feature point extraction and sliding window matching algorithm was utilized to obtain the disparity value of the detected targets,and the distance information of the obstacle was solved based on the disparity value and camera intrinsic parameters.Finally,autonomous obstacle avoidance of the UAV was achieved based on the position and distance of the obstacle.The obstacle avoidance algorithm was implemented on an embedded system,an average FPS of 14.3 was achieved,and the feasibility of the algorithm was verified through UAV flight testing.The improved network had an average detection accuracy of 76.88%,which was 0.37%lower than that of YOLOv5,but the detection time and parameter quantity were reduced by 22%and 25%,respectively.This algorithm has significant value for the autonomous obstacle avoidance of UAVs.

关 键 词：目标检测 轻量化 特征匹配 无人机避障 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]