摘要
由于传统的同步定位与建图(simultaneous localization and mapping,SLAM)中有很强的静态刚性假设,故系统定位精度和鲁棒性容易受到环境中动态对象的干扰。针对这种现象,提出一种在室内动态环境下基于深度学习的视觉SLAM算法。基于ORB-SLAM2进行改进,在SLAM前端加入多视角几何,并与YOLOv5s目标检测算法进行融合,最后对处理后的静态特征点进行帧间匹配。实验使用TUM数据集进行测试,结果显示:SLAM算法结合多视角几何、目标检测后,系统的绝对位姿估计精度在高动态环境中相较于ORB-SLAM2有明显提高。与其他SLAM算法的定位精度相比,改进算法仍有不同程度的改善。
Due to the strong static rigid assumption in the traditional simultaneous localization and mapping(SLAM),the system positioning accuracy and robustness are easily disturbed by dynamic objects in the environment.In view of this phenomenon,a visual SLAM algorithm based on deep learning in an indoor environment is proposed.Improving ORB-SLAM2,this research adds multi-view geometry to the front end of SLAM,integrates it with the YOLOv5s target detection algorithm,and finally performs frame-to-frame matching on the processed static feature points.The experiment uses the TUM data set for testing,and it is found that after combining multi-view geometry,target detection and SLAM algorithm,the absolute pose estimation accuracy of the system is significantly improved compared with ORB-SLAM2 in a highly dynamic environment.Compared with the positioning accuracy of other SLAM algorithms,this method also has different degrees of improvement.
作者
张庆永
杨旭东
ZHANG Qingyong;YANG Xudong(School of Mechanical and Automotive Engineering,Fujian University of Technology,Fuzhou 305118,China;Fujian Automotive Electronics and Electric Drive Laboratory,Fuzhou 305118,China)
出处
《贵州大学学报（自然科学版）》
2023年第6期53-61,共9页
Journal of Guizhou University:Natural Sciences
基金
福建省省级科技资助项目(GY-Z21004)
福建工程学院科研基金资助(GY-Z20170)。
