考虑几何约束与语义分割动态点的动态目标自适应筛选视觉SLAM方法

Visual SLAM Method with Adaptive Dynamic Object Filtering Considering Geometric Constraints and Dynamic Points in Semantic Segmentation

摘要:

针对动态场景下视觉同步定位与地图构建（SLAM）因几何约束受动态点干扰以及不易准确判断动态目标的实际运动状态影响定位精度的问题，本文提出一种考虑预定义和未定义动态点对几何约束影响与语义分割区域动态点密度的动态目标自适应筛选视觉SLAM方法。通过语义分割筛选并标记动态场景中预定义的动态目标，获得伪动态掩码语义信息。通过金字塔光流跟踪提取相邻帧的光流匹配点集，滤除伪动态掩码中光流点获得伪静态光流点集，采用双阈值迭代计算抑制动态点对几何约束的影响，并获得精确的动态点集。计算语义分割区域动态点密度确定预定义动态目标的实际运动状态，实现动态目标自适应筛选的视觉SLAM。最后，在TUM公共数据集和室内真实场景中验证所提方法的有效性。结果表明，与ORB-SLAM3以及DS-SLAM相比，本文方法在低动态场景序列与高动态场景序列中的绝对轨迹均方根误差平均降低约30.33%、92.17%和4.62%、42.5%;与SG-SLAM以及SSF-SLAM相比，本文方法在部分场景序列中的精度更高。

Abstract:

Aiming to address the issue where visual simultaneous localization and mapping (SLAM) in dynamic scenes suffered from degraded localization accuracy due to geometric constraints being disturbed by dynamic points and the difficulty in accurately determining the actual motion states of dynamic objects, a visual SLAM method with adaptive dynamic object filtering was proposed, which accounted for both the impact of predefined and undefined dynamic points on geometric constraints and the dynamic point density within semantic segmentation regions. Firstly, predefined dynamic objects in dynamic scenes were screened and labeled through semantic segmentation to obtain semantic information of pseudo-dynamic masks. Secondly, matching point sets between adjacent frames were extracted via pyramidal optical flow tracking. By filtering out optical flow points within the pseudo-dynamic mask, a pseudo-static optical flow point set was obtained. A dual-threshold iterative computation was further employed to mitigate the impact of dynamic points on geometric constraints, thereby yielding a refined set of dynamic points. Then, the dynamic point density within each semantic segmentation region was calculated to determine the actual motion state of predefined dynamic objects, thereby achieving visual SLAM with adaptive dynamic object filtering. Finally, the effectiveness of the proposed method was validated on the TUM public dataset and in real indoor scenes. The results showed that compared with ORB-SLAM3 and DS-SLAM, the proposed method reduced the root mean square error of the absolute trajectory by approximately 30.33% and 92.17% for low-dynamic sequences and by 4.62% and 42.5% for high-dynamic sequences, respectively. Compared with SG-SLAM and SSF-SLAM, the proposed method achieved higher accuracy in some scene sequences.