Scholarlink

Real-time SLAM for UAVs operating in dynamic environments presents significant challenges owing to interference from moving objects, degradation of feature tracking, and localization accuracy. This study introduces a YOLOv11-integrated ORB-SLAM3 framework to eliminate dynamic object features in real-time. The system leverages a TensorRT-optimized YOLOv11 detector running parallel to ORB-SLAM3's GPU-accelerated tracking thread, ensuring low-latency inference and minimal tracking disruption. Experiments conducted using Microsoft AirSim’s photorealistic urban simulation demonstrated substantial performance gains. In the long-duration, high-dynamic scenario, the proposed system reduced the Absolute Trajectory Error (ATE) RMSE from 0.0536 m (baseline) to 0.0026 m, representing a 95.2% improvement. The SLAM tracking thread sustained ~ 74 FPS, whereas the YOLOv11 detector operated asynchronously at ~ 22 FPS. These results confirm that dynamic object filtering using YOLOv11 can significantly enhance SLAM accuracy without compromising real-time performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-Time Dynamic Object Elimination in ORB-SLAM3 for UAV Navigation Using YOLOv11: A Simulation-Based Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Htet Paing Soe,
  • Liangyu Zhao,
  • Yeqing Zhu,
  • Syeda Amna Raza,
  • Alkhalil Siddick Adam,
  • Rehman Maqsood Ur

摘要

Real-time SLAM for UAVs operating in dynamic environments presents significant challenges owing to interference from moving objects, degradation of feature tracking, and localization accuracy. This study introduces a YOLOv11-integrated ORB-SLAM3 framework to eliminate dynamic object features in real-time. The system leverages a TensorRT-optimized YOLOv11 detector running parallel to ORB-SLAM3's GPU-accelerated tracking thread, ensuring low-latency inference and minimal tracking disruption. Experiments conducted using Microsoft AirSim’s photorealistic urban simulation demonstrated substantial performance gains. In the long-duration, high-dynamic scenario, the proposed system reduced the Absolute Trajectory Error (ATE) RMSE from 0.0536 m (baseline) to 0.0026 m, representing a 95.2% improvement. The SLAM tracking thread sustained ~ 74 FPS, whereas the YOLOv11 detector operated asynchronously at ~ 22 FPS. These results confirm that dynamic object filtering using YOLOv11 can significantly enhance SLAM accuracy without compromising real-time performance.