动态场景下结合语义的半直接法视觉里程计

Semi-direct visual odometer incorporating semantics in dynamic scene

针对传统视觉里程计方法在动态环境中跟踪效果差、容易失效的问题,提出了一种适用于动态场景下融合语义的半直接法视觉里程计,主要由语义分割、位姿跟踪和地图管理三部分组成。首先,使用语义分割网络Mask R-CNN对每一帧图像进行语义分割,得到语义分割二值图,结合语义先验信息删除动态特征;然后,利用静态部分特征进行帧间对齐和特征匹配完成位姿跟踪;最后,通过筛选关键帧和更新3D点深度对地图更新,并利用光束平差算法对局部地图的位姿和3D点进行优化。在公共数据集TUM RGB-D的动态序列集上的实验结果表明,该方法与ORB-SLAM2、OpenVSLAM相比,其跟踪误差分别减少了68%和72%,验证了该方法在动态场景中位姿跟踪的准确性。

To address the problem that traditional visual odometer method has poor tracking effect and easy failure in dynamic environment,this paper proposed a semi-direct visual odometer combined with semantics in dynamic scene.This method consisted of three parts:semantic segmentation,pose tracking and map management.Firstly,it used the semantic segmentation network Mask R-CNN to segment each frame of image,and obtained the binary image of semantic segmentation.It deleted the dynamic features by combining the semantic prior information.Secondly,it used the static part features for inter frame alignment and feature matching to complete pose tracking.Finally,it updated the map by selecting key frames and the depth of 3D points,and optimized the pose and 3D points of local map by using the bundle-adjustment algorithm.Compared with ORB-SLAM2 and OpenVSLAM on the dynamic sequence set of the public data set TUM RGB-D,the experimental results show that the tracking error decreases 68%and 72%,respectively.It shows the accuracy of this method in dynamic scene.