摘要
针对SLAM系统在动态场景中因物体快速移动导致特征匹配性能降低的问题,本文在ORB-SLAM2框架上提出基于改进YOLOv8s的动态视觉SLAM算法。使用轻量级网络Fasternet替换YOLOv8s主干网络,使用RT-DETR中的Transformer Decoder Head改进检测头。结合几何与语义信息实现动态特征点的高效剔除。在TUM数据集上的实验表明,本算法在动态场景下的定位与建图精度比ORB-SLAM2提高约96.06%,并且具有良好的实时性。
Aiming at the problem of feature matching performance of SLAM system is reduced due to fast moving objects in dynamic scenes,this paper proposes a dynamic visual SLAM algorithm based on improved YOLOv8s on ORB-SLAM2 framework.The YOLOv8s core network is replaced with lightweight network Fasternet,and the detection head is improved with Transformer Decoder Head in RT-DETR.Combining with geometric and semantic information,the dynamic feature points are eliminated efficiently.Experiments on TUM data set show that,the localization and mapping accuracy of this algorithm in dynamic scenes increase by about 96.06% compared with ORB-SLAM2,and it has good real-time performance.
作者
廖涛
李智
Liao Tao;Li Zhi(School of Electronics and Information Engineering,Sichuan University,Chengdu 610065,Sichuan Province,China)
出处
《科学与信息化》
2024年第10期40-42,共3页
Technology and Information
关键词
SLAM
目标检测
动态特征点剔除
定位建图精度
SLAM
target detection
dynamic feature point elimination
localization and mapping accuracy
