Visual SLAM in Dynamic Environments Based on Object Detection and Scene Flow

In Simultaneous Localization and Mapping (SLAM) algorithms, it is common to assume static environments. However, Static assumption may cause the SLAM system to drift or even crash in many practical scenarios where there are many moving objects. Combining SLAM with dynamic object estimation can significantly improve system stability in dynamic environments. In this paper, we propose a visual SLAM system for dynamic scenes that achieves precise, robust camera pose estimation and dynamic object tracking. By utilizing deep learning-based object detection and scene flow feature point tracking technologies, dynamic objects are separated from static ones. Then, dynamic and static objects arejointly optimized, resulting in 3D coordinates of feature points as well as the 6 degree-of-freedom state representation of dynamic objects in space. Through experiments, our proposed method has been shown to achieve higher accuracy and robustness in highly dynamic scenes.