DOG-SLAM: Enhancing Dynamic Visual SLAM Precision Through GMM-Based Dynamic Object Removal and ORB-Boost

摘要

Accurate mapping and navigation in dynamic environments pose substantial challenges for autonomous robotics. Traditional visual simultaneous localization and mapping (Visual SLAM) methods frequently encounter difficulties with dynamic elements, leading to data association errors and reduced localization accuracy. This study introduces DOG-SLAM, a dynamic Visual SLAM system designed to enhance performance in dynamic environments by addressing three critical aspects: more precise detection of dynamic regions, extraction of an adequate number of static feature points, and improvement of static feature point quality. DOG-SLAM incorporates a dynamic object removal module and a descriptor enhancement module. The dynamic object removal module uses a pseudosemantic segmentation strategy based on GMM, combined with RGB and depth data, for precise segmentation of dynamic regions, thus preserving a greater number of static regions. In addition, a dynamic preremoval strategy is implemented to extract a sufficient number of static feature points, maintaining feature point stability and improving localization accuracy. The descriptor enhancement module uses FeatureBooster to generate ORBboost descriptors, thereby improving the robustness and match rate of static feature points. Experimental validation on standard datasets, including TUM and BONN, demonstrated that DOG-SLAM reduced absolute trajectory error (ATE) by up to 98.38% and 99.26%, respectively, compared with its base model ORB-SLAM3. Moreover, our approach led the field in positioning accuracy compared with current state-of-the-art dynamic SLAM technologies. These results highlight the improved accuracy of DOG-SLAM in dynamic environments, indicating its potential to enhance autonomous robot mapping and navigation.