基于Mask R-CNN的动态物体剔除方法在RGB-D视觉SLAM中的应用

运行的场景中没有运动物体是大多数SLAM算法的前提,这个假设过于理想化,导致大多数视觉SLAM算法在动态环境下无法使用,因此也就限制了其在服务型机器人和自动驾驶等中的应用。提出了一种动态物体检测及剔除方法并将其整合到ORB-SLAM21]算法中,提升了其在动态场景中使用RGB-D摄像头时的稳定性。基于Mask R0CNN获得动态物体的检测和移除能力从而剔除从动态物体上提取到的ORB特征。在公共的RGB-D数据集上评估了加入动态物体剔除方法后的ORB-SLAM2系统,对比了在动态场景下和原系统的差异。改造后的系统在动态场景下的定位和建图精度提升较为明显。

Dynamic Object Segmentation Based on Mask R-CNN Apply in RGB-D SLAM

This paper presents a method that can detect dynamic object and remove it at the same time,and adds it to ORB-SLAM2 for improving the stability when use RGB-D camera in dynamic scene,capable of detecting and removing by Mask R-CNN,then can cull the ORB features that extract from dynamic object.The paper evaluates the ORB-SLAM2 system that combine with our approach on public RGB-D datasets and contrast the difference with original ORB-SLAM2 system in dynamic environment.The precision of localization and mapping that obtain from the ORB-SLAM2 system that be augmented in dynamic environment have a significant promotion.