Robust SLAM in Dynamic Scenarios Based on Deep Learning and Geometric Constraints

摘要

Visual Simultaneous Localization and Mapping (VSLAM) is critical to autonomous robotic systems. Many impressed slam systems heavily rely on a static environment assumption and face the problem of an erroneous map and wrong poses estimation caused by moving objects. In this paper, we propose a slam method based on the combination of a semantic segmentation network and geometric constraints. First, a real-time light-weight semantic segmentation network is introduced to detect the potential dynamic objects. Then a geometric module including optical flow tracking and epipolar constraints is designed to detect the dynamic points further. Hence, more effective and complete dynamic points considered as outliers are classified and the camera poses based on static points after dynamic points removal are estimated more accurately. Experiments on public datasets are conducted and the experimental results indicate that the adoption of the method proposed in this paper significantly improve the accuracy of pose estimation in highly dynamic scenarios.