激光-惯性-视觉定位与语义建图的多模态融合
Multi modal Fusion for LiDAR Inertial Visual Localization and Semantic Mapping
李泽亚 1张丹凤 1金鑫1
作者信息
- 1. 辽宁石油化工大学信息与控制工程学院，抚顺，113001
- 折叠
摘要
本文提出面向动态与感知退化环境的安全约束 LiDAR-Visual-Inertial 融合框架 TriGuard-Fusion。本框架构建于 FAST-LIVO2 的紧耦合误差状态迭代卡尔曼滤波流程之上，通过动态干扰、视觉质量、惯性创新、LiDAR 扫描到地图一致性和预更新静态置信度构建多源风险状态。SCAFC将策略输出视为候选视觉调节与有界位姿反馈动作，并通过运行可行性检查、动作投影和保守回退，限制无效候选动作进入在线估计。地图更新阶段，静态性感知接纳机制联合动态、几何、语义和时序证据，减少低可靠观测写入长期静态语义地图。在UrbanNav-HK-Medium-Urban-1 序列上，本方法取得1.44 m 的ATE RMSE，较本实验中表现最优的外部基线降低60.8%。在本文采用的代理地图评价协议下，三个自采序列中均未检测到动态残留体素，系统有效输出频率为17.05 ± 0.29 Hz。实验结果表明，TriGuard-Fusion 能够在所测试场景中兼顾定位精度、地图动态残留抑制与在线处理效率。
Abstract
This paper presents TriGuard Fusion, a safety constrained LiDAR Visual Inertial fusion framework targeting dynamic and perception degraded environments. Built on the tightly coupled error state iterative Kalman filter pipeline of FAST LIVO2, the framework constructs a multi source risk state from dynamic disturbances, visual quality, inertial innovation, LiDAR scan to map consistency and pre update static confidence. The SCAFC module treats policy outputs as candidate visual regulation and bounded pose feedback actions, and prevents invalid candidate actions from entering online estimation through operational feasibility check, action projection and conservative fallback. During the map update phase, a static awareness acceptance mechanism fuses dynamic, geometric, semantic and temporal evidence to reduce low reliability observations written into the long term static semantic map. On the UrbanNav HK Medium Urban 1 sequence, the proposed method achieves an ATE RMSE of 1.44?m, which is reduced by 60.8?% compared with the best performing external baseline in our experiments. Under the proxy map evaluation protocol adopted in this paper, no dynamic residual voxels are detected in three self collected sequences, and the effective output frequency of the system is 17.05 ± 0.29 Hz. Experimental results verify that TriGuard Fusion can simultaneously balance localization accuracy, dynamic residual suppression in maps and online processing efficiency for the tested scenarios.关键词
多传感器融合/激光-视觉-惯性 SLAM/动态环境/可靠性感知融合/静态语义建图/运行约束自适应控制Key words
autonomous driving/CARLA/driving scenarios/offline dataset/expert trajectory/data standardization引用本文复制引用
李泽亚,张丹凤,金鑫.激光-惯性-视觉定位与语义建图的多模态融合[EB/OL].(2026-09-18)[2026-09-24].http://www.paper.edu.cn/releasepaper/content/202609-29.学科分类
交通运输经济
