融合深度与语义信息的双目视觉SLAM

Binocular Vision SLAM Integrating Depth and Semantic Information

摘要: 针对动态场景下传统视觉SLAM系统因动态物体干扰导致定位精度下降的问题，本文提出一种融合深度特征与语义信息的双目视觉SLAM优化方法。首先，采用轻量化改进的DeepLabV3+语义分割网络，在保证实时性的同时提升分割精度；其次，设计语义与深度一致性融合的动态掩码构建机制：通过语义分割初筛潜在动态区域，结合双目深度图分析时序深度变化，并引入置信度加权评分及时序一致性判断策略，精准区分动态与静态特征点；最后，在ORB-SLAM3框架中集成上述模块，构建包含跟踪、局部建图与回环检测的完整系统。在Euroc和KITTI数据集上的实验表明：与ORB-SLAM3相比，本文方法在室内场景绝对轨迹误差(ATE)降低8%~15%，室外场景降低7%~35%，显著提升了动态环境下的鲁棒性与定位精度。

Abstract: This paper proposes a binocular visual SLAM optimization method that integrates depth features and semantic information, addressing the issue of reduced positioning accuracy in traditional visual SLAM systems due to interference from dynamic objects in dynamic scenes. Firstly, a lightweight improved DeepLabV3 semantic segmentation network is adopted to enhance segmentation accuracy while ensuring real-time performance; secondly, a dynamic mask construction mechanism based on semantic and depth consistency is designed: it initially filters potential dynamic areas through semantic segmentation, analyzes temporal depth changes using binocular depth maps, and introduces a confidence-weighted scoring and temporal consistency judgment strategy to accurately distinguish between dynamic and static feature points; finally, the above modules are integrated into the ORB-SLAM3 framework to construct a complete system that includes tracking, local mapping, and loop detection. Experiments on the Euroc and KITTI datasets demonstrate that compared to ORB-SLAM3, the proposed method reduces absolute trajectory error (ATE) by 8%~15% in indoor scenes and by 7%~35% in outdoor scenes, significantly enhancing robustness and positioning accuracy in dynamic environments.