基于实例分割和多约束优化的动态SLAM方法

Dynamic SLAM Based on Instance Segmentation and Multi-Constraint Optimization

摘要:

针对传统视觉SLAM方法在动态环境中易受运动目标和光照变化干扰导致定位精度和系统鲁棒性下降的问题，提出一种基于实例分割与多约束优化的动态视觉SLAM方法，通过改进的轻量化YOLOv11模型实现高效实例分割，减少动态特征对特征提取与数据关联的影响；并在特征匹配中引入局部亮度归一化和光照不变性约束，提升特征匹配精度；同时结合信息论与协方差约束设计一种关键帧选择与边缘化策略，通过量化候选关键帧对系统可观性的贡献，结合不确定性度量实现关键帧优化，在多个动态数据集和实际动态场景上的实验结果表明，该方法在高动态场景下较ORB-SLAM3的绝对轨迹误差的均方根误差和平均误差分别降低了96.19%和95.23%，显著提升了室内动态场景下的定位精度与鲁棒性。

Abstract:

Aiming at the problem that traditional visual Simultaneous Localization and Mapping (SLAM) methods are susceptible to interference from moving objects and illumination changes in dynamic environments, leading to the degradation of positioning accuracy and system robustness, a dynamic visual SLAM method based on instance segmentation and multi-constraint optimization is proposed. An improved lightweight YOLOv11 model is adopted to achieve efficient instance segmentation, which reduces the impact of dynamic features on feature extraction and data association. In addition, local brightness normalization and illumination invariance constraints are introduced into feature matching to improve the accuracy of feature matching. Meanwhile, a keyframe selection and marginalization strategy is designed by combining information theory with covariance constraints. By quantifying the contribution of candidate keyframes to system observability and combining uncertainty measurement, keyframe optimization is realized. Experimental results on multiple dynamic datasets and real dynamic scenes show that, compared with ORB-SLAM3 in high-dynamic scenes, the proposed method reduces the root mean square error and average error of the absolute trajectory error by 96.19% and 95.23% respectively, which significantly improves the positioning accuracy and robustness in indoor dynamic environments.