Visual SLAM based on instance segmentation in dynamic scenes

There is a problem in that the existing visual simultaneous localization and mapping (SLAM) algorithm mainly applied to static scenes is less likely to be applied to dynamic scenes directly. A dynamic environment SLAM algorithm is illustrated in this paper based on instance segmentation and epipolar geometry to narrow down the interference between moving objects and localization accuracy of SLAM. For the reason of semantic information, the feature points on moving objects are removed. In terms of the dynamic region without prior knowledge, deletion of the dynamic points can be achieved by the epipolar geometry method to reduce the impact of dynamic points on positioning accuracy. On the premise of that, the experimental verification on the TUM public dataset and comparison with other classical algorithms is done on the proposed algorithm. The results show that effective detection can be realized to remove potential and uncertain moving objects. The SLAM system effectively enhances the robustness of SLAM in highly dynamic scenes and significantly improves the localization accuracy.