Indoor Dynamic Scene Visual SLAM Based On Human Detection And Geometric Constraints

The existence of dynamic objects dramatically depresses the performance of localization and mapping of SLAM systems. In this paper, we propose a novel indoor dynamic visual SLAM framework based on human detection and geometric constraints to solve the problem caused by the dynamic object. We first use semantic segmentation model, DeeplabV3, to detect human and then use spatial geometric constraints to further remove dynamic feature points. Several experiments on TUM datasets and Bonn datasets have shown that our performances, in terms of absolute trajectory error (ATE), are 91% to 99% higher than that of ORBSLAM2.