面向室内动态场景的语义感知与点线融合视觉SLAM系统

A Semantic-Aware Visual SLAM System with Point-Line Feature Fusion for Indoor Dynamic Environments

摘要:

针对室内动态场景下传统视觉SLAM因环境静态假设失效引发的定位精度下降、鲁棒性差等问题，设计一种RGB-D相机下结合目标检测与点线融合的SLAM系统（YEL-SLAM）；采用集成自适应长度抑制与角度过滤的ELSED线特征提取策略，在ORB点特征基础上引入线特征作为几何维度的补充；融合YOLOv11n目标检测模型与深度残差一致性判定，构建语义-几何联合的动态特征筛选机制，实现对动态点线特征的精准判别与滤除；并建立点线权重自适应调节的联合优化模型；TUM数据集实验测试显示，相较于ORB-SLAM3，YEL-SLAM绝对轨迹误差最大降幅达96%；与其他语义SLAM方法相比，YEL-SLAM在Walking_xyz等高动态序列上也表现出更好的鲁棒性和精度。

Abstract:

Traditional visual SLAM systems suffer from decreased localization accuracy and poor robustness in indoor dynamic scenes because the static environment assumption fails. To address these problems, this paper designs a YOLO and ELSED-based SLAM (YEL-SLAM) system for RGB-D cameras. The system combines object detection with point-line feature fusion, adopting the ELSED extraction strategy to integrate adaptive length suppression and angle filtering, which introduces line features as a geometric supplement to ORB point features. Furthermore, the system fuses the YOLOv11n object detection model with depth residual consistency checks to construct a joint semantic-geometric dynamic feature screening mechanism, achieving precise identification and removal of dynamic point and line features. Additionally, a joint optimization model with adaptive weight adjustment for points and lines is established. Experimental tests on the TUM dataset show that compared to ORB-SLAM3, YEL-SLAM significantly improves performance, with a maximum reduction in absolute trajectory error reaching 96%. Moreover, compared with other semantic SLAM methods, YEL-SLAM shows better robustness and accuracy in highly dynamic sequences like Walking_xyz.