动态场景下基于视觉传感器的SLAM研究

Research on SLAM Based on Vision Sensor in Dynamic Scene

针对基于视觉传感器的同步定位与地图构建(SLAM)在动态场景下鲁棒性差、定位与建图精度易受动态物体干扰的问题,提出一种动态场景下基于语义分割和几何信息的视觉SLAM算法。该算法基于ORB-SLAM2,首先,采用GCNv2进行特征提取;然后,基于语义分割网络Deeplabv3+,提出像素级的图像语义分割方法,分割潜在的先验动态目标对象;之后,设计多视图几何的动态点检测方法检测动态对象的运动状态信息,并将剩余的静态特征用于位姿估计;最后,结合关键帧和深度图,生成包含语义信息的点云地图和八叉树地图。该算法在TUM数据集上进行了对比验证分析,结果表明该算法有效提高了VSLAM系统在动态场景下的性能。

Aiming at the problems that simultaneous localization and map construction(SLAM)based on vision sensor is poor in robustness and easy to be interfered by dynamic objects in dynamic scenes,a visual SLAM method based on semantic segmentation and geometric information in dynamic scenes is proposed.This method is based on ORB-SLAM2.Firstly,GCNv2 is used for feature extraction,and then based on the semantic segmentation network Deeplabv3+,a pixel level image semantic segmentation method is proposed to segment potential prior dynamic target objects,and then a multi-view geometry dynamic point detection method is designed to detect the motion state information of dynamic objects.The remaining static features are used for pose estimation.Finally,the point cloud map and octree map containing semantic information are generated by combining the key frame and depth map.The proposed method is compared and verified on TUM data set,and the results show that the proposed algorithm can effectively improve the performance of VSLAM system in dynamic scenarios.