福州大学学报（自然科学版）2026，Vol.54Issue(2)：154-161,8.DOI:10.7631/issn.1000-2243.25069
动态场景下基于语义分割的双目惯性SLAM
Astereo inertial SLAM based on semantic segmentation for dynamic environment
摘要
Abstract
A semantic segmentation-based stereo inertial simultaneous localization and mapping(SLAM)method is proposed for dynamic scenes to address the impact of dynamic objects on SLAM using a stereo direct method.The frontend employs adaptive optical flow thresholding for dynamic feature recognition,filtering out truly moving features to enhance localization accuracy in dynamic environments.During coarse tracking,it fuses static stereo residual constraints(after dynamic object removal)with inertial measurement unit data to correct the visual pose,addressing the low accuracy and poor robustness of stereo direct visual odometry.For backend optimization,a sliding window approach tightly couples stereo and inertial measurement unit(IMU)data.A novel cross-frame bidirec-tional feature projection method constructs projection residuals from the host frame's left image to the target frame's right image,ensuring stable pose estimation even in scenes with abundant dynamic features.Evaluations on the KITTI dataset and a custom dataset demonstrate that the improved dynamic detection effectively filters dynamic features.Compared to stereo direct SLAM,the proposed method achieves improvements of 23.89%in translational accuracy and 20.18%in rotational accuracy,while reducing the root mean square error of the motion trajectory by 33.14%on the custom dataset.关键词
动态环境/光流法/双目直接法/IMU融合Key words
dynamic environment/optical flow method/stereo direct method/IMU fusion分类
信息技术与安全科学引用本文复制引用
林申炀,彭育辉,孙宝哲,张淦,张家铭..动态场景下基于语义分割的双目惯性SLAM[J].福州大学学报（自然科学版）,2026,54(2):154-161,8.基金项目
福建省科技厅引导性基金资助项目(2022H0007) （2022H0007）
