Scholarlink

<p>This article proposes a real-time SLAM framework that integrates pixel level, object level, and scene level semantic information to address the issue of traditional visual SLAM being susceptible to cumulative errors and semantic blind spots in dynamic and complex environments. This framework collaboratively optimizes trajectory optimization and error correction through multi-scale semantic constraints. The system constructs a lightweight multi branch semantic segmentation network to generate a unified semantic graph, and embeds it into the entire process of front-end tracking, keyframe selection, and back-end pose graph optimization, achieving semantic driven trajectory generation and residual feedback correction. Experiments on the TUM RGB-D, KITTI, and self collected Urban Campus-2025 datasets show that the proposed method significantly outperforms existing mainstream systems in dynamic interference and repetitive structure scenarios: ATE (absolute trajectory error) is as low as 0.112&#xa0;m, RPE (relative pose error) is 0.021&#xa0;m, semantic annotation IoU reaches 78.4%, and single frame processing delay is controlled at 35.2 milliseconds. Especially during long-term operation, the system effectively suppresses drift through semantic closed-loop, and the final ATE of the 45&#xa0;min sequence is only 0.103&#xa0;m. The results validated the effectiveness of multi-scale semantic constraints in improving localization accuracy, trajectory rationality, and system robustness, providing a feasible technical path for highly reliable visual navigation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time Visual SLAM Trajectory Optimization and Error Correction Technology under Multi-scale Semantic Constraints

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月16日

查看全文
  • Xiaohu Liu,
  • Naigen Li

摘要

This article proposes a real-time SLAM framework that integrates pixel level, object level, and scene level semantic information to address the issue of traditional visual SLAM being susceptible to cumulative errors and semantic blind spots in dynamic and complex environments. This framework collaboratively optimizes trajectory optimization and error correction through multi-scale semantic constraints. The system constructs a lightweight multi branch semantic segmentation network to generate a unified semantic graph, and embeds it into the entire process of front-end tracking, keyframe selection, and back-end pose graph optimization, achieving semantic driven trajectory generation and residual feedback correction. Experiments on the TUM RGB-D, KITTI, and self collected Urban Campus-2025 datasets show that the proposed method significantly outperforms existing mainstream systems in dynamic interference and repetitive structure scenarios: ATE (absolute trajectory error) is as low as 0.112 m, RPE (relative pose error) is 0.021 m, semantic annotation IoU reaches 78.4%, and single frame processing delay is controlled at 35.2 milliseconds. Especially during long-term operation, the system effectively suppresses drift through semantic closed-loop, and the final ATE of the 45 min sequence is only 0.103 m. The results validated the effectiveness of multi-scale semantic constraints in improving localization accuracy, trajectory rationality, and system robustness, providing a feasible technical path for highly reliable visual navigation.