A Visual SLAM Based on Dynamic Object Removal for Small-scale Robotic Rat Navigation

Environmental perception is the foundation of navigation and plays an important role in the practical applications of robots. Visual SLAMs commonly suffer from erroneous data association due to moving objects in the real world. To mitigate the issue of low accuracy, we present a robust visual SLAM method based on dynamic object removal for robot navigation. The method creates criteria for determining dynamic properties based on objects' semantic and depth information. Besides, an iteratively updated dynamic object removal method is included in the SLAM framework to optimize the overall localization accuracy. At last, we performed evaluations on dynamic benchmark datasets to demonstrate the competitiveness of the proposed method, reducing absolute trajectory error by 86.12% compared to ORB-SLAM3. In addition, real-world SLAM and navigation experiments on a robotic rat were also conducted and the results proved that the proposed method outperforms state-of-the-art methods with an accuracy of over 97.48%.