动态环境下基于改进YOLOv5s目标检测的视觉SLAM方法

打开文本图片集

中图分类号：TP391

文献标识码：A

Abstract：Traditional visual SLAM systems suffer from three critical defects in dynamic environments：illumination sensitivity leading to mapping distortion，moving objects causing pose drift， and trajectory artifacts from dynamic objects contaminating the map topology，which severely degrade both system robustness and map authenticity. To address these issues，this paper introduces an improved YOLOv5s model embedded into a visual SLAM framework， proposing a dynamic visual SLAM method based on enhanced YOLOv5s object detection-named DYFO-SLAM （Dynamic YOLO-Flow Odometry SLAM）. The method incorporates optical flow and epipolar constraints to eliminate dynamic features. Evaluations on public datasets and real-world scenarios demonstrate that our approach effectively removes dynamic features， significantly improving the performance of traditional visual SLAM systems in dynamic environments.

Keywords： visual SLAM;deep learning；optical flow;epipolar constraint

传统视觉同步定位与地图构建（Simultaneous Localization and Mapping，SLAM）系统通常基于静态环境假设。（剩余8288字）