Research and Application of SLAM Algorithm for Mobile Robots in Indoor Dynamic Scene

In this paper, the fusion object detection SLAM algorithm based on ORB-SLAM3 was proposed to improve the positioning accuracy of indoor dynamic scene and realize real-time operation on embedded mobile terminal. In order to transfer the image input from the visual odometer to the object detection thread, the YOLOv8 object detection thread was added to the ORB-SLAM3 visual SLAM algorithm. Then the dynamic target marking was implemented, and feature points were extracted from outside the dynamic object area to achieve feature extraction and avoid dynamic objects interfering with map construction. The proposed algorithm's real-time performance in the embedded mobile terminal can be guaranteed because TensorRT technology was used to accelerate the object detection thread. Then the object detection time of each frame measured in the embedded mobile terminal was approximately 20 ms. It was demonstrated that the enhanced algorithm performed well in the experiments after it was validated on the public TUM data set. Comparing the before improvement result of the ORB-SLAM3 algorithm with the new result, the positioning accuracy could be increased by 95% in indoor high dynamic scene and slightly in low dynamic scene.