基于单目视觉与改进势场法的避障算法  

Obstacle avoidance algorithm based on monocular vision and improved potential field

作　　者：纪宇航 蔡文靖 刘鑫 王礼贺 JI Yu-hang;CAI Wen-jing;LIU Xin;WANG Li-he(CETC Electro-Optics Technology Co.Ltd.,Beijing 100015,China)

机构地区：[1]中电科光电科技有限公司,北京100015

出　　处：《激光与红外》2024年第6期980-990,共11页Laser & Infrared

摘　　要：针对无人平台在未知环境缺乏全局障碍物位置信息难以进行实时避障的问题,提出一种融合深度神经网络和改进人工势场的动态避障方法。首先利用YOLOv5s轻量目标检测网络和轻量化深度估计网络构建障碍感知模块,探测障碍物位置和深度;其次使用目标框和等效深度描述周围障碍物的三维信息;然后将自身平台投影到图像平面形成核心区,并根据障碍物的等效深度网格与核心区的位置关系计算核心区在像平面虚拟势场受力方向、所需要的偏航角和线速度;最后控制系统接收到信号后引导无人平台转向或制动,使核心区内部深度大于设定的安全距离完成动态避障。实验基于单目可见光和红外进行避障测试,结果表明视觉感知模块可以准确探测常见障碍物的位置和深度,核心区能够直观地反映无人平台与障碍物的位置关系,方法仅依赖单目传感器便可实现有效避障,相比于传统方法成本低廉、部署灵活,为无人平台在日间、夜间的未知环境进行避障提供了新思路。In unknown environments without global obstacle location information,real time avoidance is a challenging task for unmanned platforms.To address this issue,a method that fuses deep neural networks(DNNs)with an improved artificial potential field(APF)algorithm is proposed in this paper.Firstly,YOLOv5s and a lightweight depth estimation model is used to construct an obstacle perception module to detect the location and depth of obstacles.Then,the target frame and equivalent depth are utilized to describe the three dimensional information of the surrounding obstacles.Subsequently,the platform is projected onto the image plane,and the core area is calculated based on the positional relationship between the obstacle's equivalent depth grid and the core area to compute the direction of the core area in terms of the force on the virtual potential field in the image plane,the required yaw angle,and the linear velocity.Finally,the control system receives the signals and guides the unmanned platform to steer or brake,ensuring that the internal depth of the core area is greater than the safety distance.The experiment is based on monocular visible light and infrared for obstacle avoidance test,and the results demonstrate that the perception module can accurately detect the location and depth of common obstacles and the core area can intuitively reflect the positional relationship between the unmanned platform and surrounding obstacles.Compared to traditional methods,the proposed method relies on a monocular sensor alone to effectively avoid obstacles,achieving lower cost and flexible deployment,which provides a new idea for unmanned platforms to avoid obstacles in unknown environments during daytime and nighttime.

关 键 词：深度学习 动态避障 人工势场 

分 类 号：TN29[电子电信—物理电子学] TP274[自动化与计算机技术—检测技术与自动化装置]