ABSTRACT Traditional simultaneous localization and mapping (SLAM) systems exhibit strong performance in static scenarios, while they continue to face significant difficulties in dynamic scenarios characterized by moving objects and sparse features, which adversely affect the robustness and accuracy of the systems. Current semantic dynamic visual SLAM approaches commonly incorporate either semantic segmentation or object detection techniques. However, the former tends to be computationally demanding, while the latter is often susceptible to errors and omissions in detection. This paper proposes a novel visual‐inertial SLAM system designed for Dynamic environments integrating YOLOv5s‐Asym and Scene flow, referred to as DAS‐SLAM. Within the DAS‐SLAM framework, a novel object detection method incorporating asymmetric module has been developed, resulting in enhanced accuracy in object detection. Furthermore, DAS‐SLAM integrates the proposed YOLOv5s‐Asym network with geometric constraints derived from scene flow to effectively remove dynamic feature points, thereby substantially improving the accuracy of the system. Extensive indoor and outdoor experiments indicate that the proposed approach exhibits enhanced accuracy and robustness in comparison to leading contemporary methods.
潘 et al. (Sun,) studied this question.