针对传统语义实时定位与建图(Semantic real-time Localization and Mappling，SLAM)算法在动态环境下剔除特征点过多、造成定位精度降低的问题，提出一种基于实例分割与光流的视觉语义SLAM算法。算法使用Mask R-CNN网络对图像中的潜在动态物体进行实例级别的分割，同时在光流线程中对动态物体进行识别并剔除，随后使用剩余的静态光流点与静态特征点联合优化定位，实现语义信息与光流信息的充分融合利用。使用公开数据集测试和地面无人平台实验对所提方法进行验证。实验结果表明，在TUM数据集下，新方法的定位均值误差相比ORB-SLAM2 平均提高75%，相比Dyna-SLAM平均提高8。5%。
A SLAM in Dynamic Environment Based on Instance Segmentation and Optical Flow
A visual semantic SLAM algorithm based on instance segmentation and optical flow is proposed to address the issue of excessive removal of features by traditional semantic SLAM algorithms in dynamic environments.The proposed algorithm utilizes a Mask R-CNN network to perform the instance-level segmentation of potential dynamic objects in an image,and also identifies and eliminates dynamic objects in the optical flow thread.The remaining static optical flow points and static feature points are then used to optimize the location estimation process,ensuring the optimal utilization of both semantic and optical flow information.The proposed algorithm is validated through testing on open datasets and an unmanned ground platform experiment.The experimental results indicate that the average error of the proposed algorithm is 75%and 8.5%lower than those of ORB-SLAM2 and Dyna-SLAM,respectively,on TUM dataset.
dynamic environmentoptical flow methodinstance segmentationsemantic simultaneous localization and mapping
万方数据
维普
