RESEARCH ON SEMANTIC-ASSISTED SLAM IN COMPLEX DYNAMIC INDOOR ENVIRONMENT
Keywords: Simultaneous localization and mapping, Semantic segmentation, Dynamic object recognition, Robot indoor navigation, Scene understanding
Abstract. Visual Simultaneous Localization and Mapping (SLAM) systems have been widely investigated in response to requirements, since the traditional positioning technology, such as Global Navigation Satellite System (GNSS), cannot accomplish tasks in restricted environments. However, traditional SLAM methods which are mostly based on point feature tracking, usually fail in harsh environments. Previous works have proven that insufficient feature points caused by missing textures, feature mismatches caused by too fast camera movements, and abrupt illumination changes will eventually cause state estimation to fail. And meanwhile, pedestrians are unavoidable, which introduces fake feature associations, thus violating the strict assumption that the unknown environment is static in SLAM. In order to ensure how our system copes with the huge challenges brought by these factors in a complex indoor environment, this paper proposes a semantic-assisted Visual Inertial Odometer (VIO) system towards low-textured scenes and highly dynamic environments. The trained U-net will be used to detect moving objects. Then all feature points in the dynamic object area need to be eliminated, so as to avoid moving objects to participate in the pose solution process and improve robustness in dynamic environments. Finally, the constraints of inertial measurement unit (IMU) are added for low-textured environments. To evaluate the performance of the proposed method, experiments were conducted on the EuRoC and TUM public dataset, and the results demonstrate that the performance of our approach is robust in complex indoor environments.