融合语义信息的视觉惯性SLAM算法  

Visual inertial SLAM algorithm with semantic information fusion

作　　者：何铭臻 何元烈[1] 胡涛 He Mingzhen;He Yuanlie;Hu Tao(School of Computer,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广州510006

出　　处：《计算机应用研究》2024年第8期2533-2539,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(62102097)。

摘　　要：针对传统SLAM算法在动态环境中会受到动态特征点的影响,导致算法定位精度下降的问题,提出了一种融合语义信息的视觉惯性SLAM算法SF-VINS(visual inertial navigation system based on semantics fusion)。首先基于VINS-Mono算法框架,将语义分割网络PP-LiteSeg集成到系统前端,并根据语义分割结果去除动态特征点;其次,在后端利用像素语义概率构建语义概率误差约束项,并使用特征点自适应权重,提出了新的BA代价函数和相机外参优化策略,提高了状态估计的准确度;最后,为验证该算法的有效性,在VIODE和NTU VIRAL数据集上进行实验。实验结果表明,与目前先进的视觉惯性SLAM算法相比,该算法在动态场景和静态场景的定位精度和鲁棒性均有一定优势。Aiming at the problem that the positioning accuracy of traditional SLAM algorithm decreases due to the influence of dynamic feature points in dynamic environment,this paper proposed SF-VINS which fused semantic information.Firstly,it incorporated the PP-LiteSeg semantic segmentation network into the front-end of system based on the VINS-Mono algorithm and removed the dynamic feature points according to the results of semantic segmentation.Then,it constructed the semantic probability error constraint with the pixel semantic probabilities and employed adaptive weights for feature points in the backend.Based on these modifications,this paper proposed a new BA cost function and an optimization strategy for camera extrinsic para-meters to improve the accuracy of state estimation.Finally,this paper conducted experiments on the VIODE and NTU VIRAL datasets to validate its effectiveness.The experimental results show that the proposed algorithm has advantages in terms of positioning accuracy and robustness in both dynamic and static scenes compared with state-of-the-art visual-inertial SLAM algorithms.

关 键 词：动态物体 语义概率 位姿估计 视觉惯性SLAM 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]