%0 Journal Article %T 动态场景下基于YOLO的场景重建
Scene Reconstruction Based on YOLO in Dynamic Environments %A 沈康鹏 %J Modeling and Simulation %P 179-185 %@ 2324-870X %D 2025 %I Hans Publishing %R 10.12677/mos.2025.143213 %X 在实际动态环境中，深度传感器在获取环境信息时不可避免地会受到运动物体的干扰。如何有效处理动态物体、使机器人准确理解周围环境并完成复杂任务，仍然是一个亟待解决的难题。本文提出了一种基于改进型ORB-SLAM3与改进的YOLOv5相结合的语义分割方法。该方法通过识别并剔除动态特征，同时最大程度地保留静态环境的有效特征，结合ORB-SLAM3算法实现了高精度的场景重建，成功生成稠密点云地图。实验结果表明，在TUM-RGB-D数据集上，本文提出的方法相比原始ORB-SLAM3算法，在高动态场景中的RMSE平均降低了92.04%，在低动态场景中的RMSE平均降低了19.48%。特别是在动态物体比例较高的场景中，系统表现出优异的鲁棒性和准确性。此外，本文还对系统的实时性进行了优化，通过轻量化的目标检测网络和高效的特征筛选策略，确保了系统在普通硬件平台上的实时运行能力。研究结果为解决动态环境下的视觉SLAM问题提供了一种高效可靠的解决方案。
In real-world dynamic environments, depth sensors inevitably encounter interference from moving objects while acquiring environmental information. How to effectively process dynamic objects, enable robots to accurately understand their surroundings, and accomplish complex tasks remains a challenging problem. This paper proposes a semantic segmentation method that combines an improved ORB-SLAM3 with an enhanced YOLOv5. The method identifies and eliminates dynamic features while maximally preserving effective features of the static environment. By integrating the ORB-SLAM3 algorithm, it achieves high-precision scene reconstruction and successfully generates dense point cloud maps. Experimental results on the TUM-RGB-D dataset show that compared to the original ORB-SLAM3 algorithm, our proposed method reduces RMSE by an average of 92.04% in highly dynamic scenes and 19.48% in low dynamic scenes. The system demonstrates excellent robustness and accuracy, particularly in scenarios with a high proportion of dynamic objects. Additionally, we optimized the system’s real-time performance through a lightweight object detection network and efficient feature filtering strategy, ensuring real-time operation on standard hardware platforms. The research provides an efficient and reliable solution for visual SLAM problems in dynamic environments. %K 同时定位和建图， %K 动态场景， %K ORBSLAM3， %K 语义分割
Simultaneous Localization and Mapping (SLAM) %K Dynamic Scenes %K ORB-SLAM3 %K Semantic Segmentation %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=109386