一种基于目标检测的动态环境下视觉定位系统  

A visual localization system based on object detection in dynamic environment

作　　者：钟兴军[1] 吴俊琦[1] ZHONG Xingjun;WU Junqi(Student Innovation Center,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学学生创新中心,上海200240

出　　处：《现代电子技术》2024年第2期160-164,共5页Modern Electronics Technique

摘　　要：传统的基于同时定位与建图模型的视觉定位方法需要满足目标点静止假设,但大多数小型机器人的实际应用场景为动态,这限制了现有视觉定位算法在小型机器人上的使用。为此,文中使用YOLOv5卷积神经网络对环境中的动态目标进行检测,然后剔除分布在图中的移动特征点,进而改进位姿估计准确性的动态消除方法,并将此方法集成于ORBSLAM2视觉定位系统。改进方案在TUM公共动态数据集上的测试结果表明,基于YOLOv5的检测方法能够快速、准确地识别场景中的动态目标,并显著降低动态环境下位姿估计的绝对误差和相对漂移,是一种有效的动态场景视觉定位方案。The traditional visual localization method based on simultaneous localization and mapping model needs to meet the static assumption of the target point,but the actual application scenario of most small robots is dynamic,which limits the use of existing visual positioning algorithms on small robots.Therefore,a dynamic elimination method is proposed,which can use YOLOv5 convolutional neural network to detect dynamic targets in the environment,and remove the moving feature points distributed in the graph,to improve the accuracy of pose estimation.This method is integrated into the ORB‐SLAM2 visual positioning system.The testing results of the improved method on the TUM public dynamic dataset show that the YOLOv5‐based detection method can quickly and accurately identify the dynamic target in the scene,and significantly reduce the absolute error and relative drift of pose estimation in dynamic environments.It is an effective visual positioning scheme for dynamic scenes.

关 键 词：视觉SLAM 目标检测 定位系统 YOLOv5 特征点提取 动态消除 

分 类 号：TN911.23‐34[电子电信—通信与信息系统] TP242.6[电子电信—信息与通信工程]