非结构化环境下的单目视觉可通行区域检测  

Accessible Space Detection Based on Monocular Vision in Unstructured Environment

作　　者：郭植星 曾碧[1] 刘建圻[1] 陈文轩 王俊丰 GUO Zhixing;ZENG Bi;LIU Jianqi;CHEN Wenxuan;WANG Junfeng(School of Computers,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广州510006

出　　处：《计算机工程与应用》2022年第24期298-306,共9页Computer Engineering and Applications

基　　金：广东省自然科学基金(2018A030310540,2018A030313868);清远市工业高新技术领域技术攻关项目(2020KJJH039)。

摘　　要：目前三维避障主要采用三维激光雷达或者基于深度学习的障碍物识别,但前者价格昂贵,后者训练成本高且不稳定。为了稳定、鲁棒地实现低成本三维空间避障方案,提出了一种基于单目相机的可通行区域检测方法。该方法利用特征点标识障碍物,通过对相机高度和旋转平面加以约束,解决单目SLAM中的尺度不一致问题,并设计了障碍物距离求解器和代价求解器对小车前方区域特征点进行处理,计算出视觉代价地图,划分可通行区域。该方法优势在于仅需要低成本的单目相机便可完成可通行区域检测任务,可方便地移植到轻量级的移动设备,计算得到的视觉代价地图亦有利于小车后续的路径规划任务。在KITTI数据集上进行的实验表明,该方法的平均运算速度能达到20 frame/s,能满足小车实时避障的要求,对于单目SLAM的尺度恢复误差为2.5%~4.9%。Three-dimensional obstacle avoidance mainly adopts 3-D laser radar or deep learning-based obstacle recognition,but the former is expensive,while the latter training cost is high and unstable.In order to realize low-cost three-dimensional space obstacle avoidance scheme,an accessible space detection method based on monocular camera is proposed.The method identifies obstacles by feature points.To solve the scale inconsistency in the monocular SLAM,the proposed method imposes certain constraints on camera height and rotation plane.Obstacle distance solver and cost solver are designed to calculate visual cost map and distinguish accessible space using the feature points of the front space.The proposed method only requires a low-cost monocular camera to complete the accessible space detection task,and can easily transplant to lightweight mobile devices.The calculated visual cost map is also conducive to the subsequent path planning task.Experiments on the KITTI dataset show that the average operation speed of the proposed method can reach 20 frame/s,which meets the requirements of real-time obstacle avoidance for the robot.The scale recovery error for monocular SLAM is 2.5%~4.9%.

关 键 词：非结构化环境 单目视觉 视觉代价地图 可通行区域 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]