FEATURE-ENHANCED DYNAMIC VISUAL SLAM WITH CHANNEL-SPACE JOINT MODELLING
International Journal of Mechatronics and Applied Mechanics · 2025
Addressing the issues of low pose estimation accuracy and poor robustness in simultaneous localisation and mapping (SLAM) in dynamic environments, this paper proposes a visual SLAM system called PDAF-SLAM, which combines dynamic target detection with the LK optical flow method.The system incorporates a lightweight attention enhancement module, Parallel Dual-Attention Fusion (PDAF), into the YOLOv8 detection network.By parallel fusion of channel attention and spatial attention, PDAF effectively improves detection accuracy and robustness.In the dynamic target detection thread, the system uses an improved YOLOv8 to perform dynamic object detection on images captured by the camera.It then combines the LK optical flow method to analyse the motion consistency of feature points within the candidate regions, thereby eliminating real dynamic points and using the remaining static feature points for pose estimation and mapping.Finally, this model is integrated into the ORB-SLAM3 framework and evaluated on the TUM dynamic dataset.Experimental results demonstrate that the PDAF-SLAM system achieves higher positioning accuracy and robustness than the original ORB-SLAM3 in dynamic scenes, validating the effectiveness and practicality of this method.