1 Northwest Institute of Mechanical and Electrical Engineering, Xianyang 712099, China
2 School of Artificial Intelligence, Xi’an University of Electronic Science and Technology, Xi’an 710126, China
Show less
文章历史+
Received
Published
Revised
2025-10-14
2026-09-08
2026-01-19
Online Date
2026-09-18
PDF (20337K)
摘要
针对动态变化环境中同步定位与地图构建(simultaneous localization and mapping,SLAM)技术因动态物体干扰导致地图可靠性不足的问题,提出一种基于动态语义特征的轻量化地图构建算法,其核心创新在于深度融合语义感知能力与八叉树概率更新机制,实现动态环境下高精度与轻量化的统一。算法首先设计一种基于区域提议网络(region proposal network,RPN)对RGB-D图像进行动态目标识别与掩膜提取,进而通过3D点云反投影与二维掩膜的交并比(intersection over union,IoU)匹配,为点云团赋予语义标签;在此基础上,创新性地将语义信息嵌入八叉树结构的多分辨率建模过程,通过语义引导的观测概率动态更新策略,对动态目标区域实施概率抑制,并对静态结构进行增强保持,从而构建出兼具几何完整性与语义理解能力的轻量化动态地图。在TUM动态场景数据集上的实验结果表明,与DenseSLAM构建的稠密点云地图、ElasticFusion构建的半稠密地图等现有先进建图方法相比,本文算法在地图点数量和内存占用指标上最小提升幅度分别为89.68%和84.89%。将二进制的轻量化语义地图集成至SLAM系统后,绝对误差均值(mean absolute percentage error,APE-Mean)和绝对误差均方根(absolute percentage error-root mean square error,APE-RMSE)分别提升16.49%和16.26%以上,证明了本文算法所构建的轻量化语义地图能在动态变化的场景下保持高度的适应性和准确性。
Abstract
To address the issue of insufficient map reliability in simultaneous localization and mapping (SLAM) caused by dynamic object interference in changing environments, a lightweight map construction algorithm was proposed based on dynamic semantic features. Its core innovation lies in deeply integrating semantic perception capabilities with an octree probability update mechanism, achieving a balance between high precision and lightweight performance in dynamic environments. Firstly, a method based on the region proposal network (RPN) was designed to perform dynamic target recognition and mask extraction on RGB-D images. Then, through 3D point cloud back-projection and the intersection over union (IoU) matching of the two-dimensional masks, semantic labels were assigned to the point cloud clusters. On this basis, an innovative approach was adopted to embed semantic information into the multi-resolution modeling process of the octree structure. Through a dynamic update strategy of observation probability guided by semantics, the dynamic target regions were probabilistically suppressed and the static structures were enhanced and maintained, thereby constructing a lightweight dynamic map that combines geometric integrity and semantic understanding capabilities. Experimental results on the TUM dynamic scene dataset show that, compared to state-of-the-art mapping methods like the dense point cloud maps built by DenseSLAM and the semi-dense maps built by ElasticFusion, the proposed method reduces the number of map points and memory usage by at least 89.68% and 84.89%, respectively. After integrating the binary lightweight semantic map into the SLAM system, the mean absolute percentage error(APE-Mean) and the absolute percentage error-root mean square error(APE-RMSE) are improved by over 16.49% and 16.26%, respectively, demonstrating that the proposed algorithm can maintain high adaptability and accuracy in dynamically changing scenes.
WANG Bo, LI Dong, DAI Chang-hua.
Lightweight Map Construction Algorithm Based on Dynamic Semantic Features[J].
Science Technology and Engineering, 2026, 26(25): 10954-10966 DOI:10.12404/j.issn.1671-1815.2507201
同步定位与地图构建(simultaneous localization and mapping,SLAM)技术[1]是机器人实现自主感知与导航[2]的基石,其目标是在未知环境中实时估计自身位姿[3]并构建环境地图。在静态、可控的实验室环境中,传统SLAM技术已趋于成熟。然而,在实际应用场景中,如室内服务、智能仓储、人机共融等环境中,动态物体(如行走的人员、移动的设备)的广泛存在对SLAM系统的可靠性与实用性提出了严峻挑战。动态物体的运动会导致相机观测数据中包含大量异常值,若直接用于位姿估计与地图更新,将引入显著误差,造成地图杂影、扭曲乃至定位跟踪失败。因此,研究能够有效处理动态环境、生成稳定可靠地图的SLAM方法,已成为推动机器人走向实际复杂应用的关键。
步骤2 3D目标语义融合。为实现从目标的2D语义信息到3D空间几何的映射,首先对原始点云进行滤波和分割处理,生成独立点云团;随后,通过相机参数将点云反投影至二维图像平面,与RPN分割模型输出的检测掩膜进行交并比(intersection over union,IoU)匹配,筛选出高重叠区域;最终为每个点云团分配语义标签,并依据其3D位置和尺寸信息构建具有语义标注的局部点云地图。
步骤1 对深度相机采集到的原始点云数据进行预处理。为提高数据质量,采用双边滤波算法对点云进行平滑去噪,该算法在保留边缘特征的同时有效抑制离群点,其核心思想是综合考虑空间邻近度与深度相似性。滤波后,采用基于欧几里得距离的具有噪声的基于密度的空间聚类算法(density-based spatial clustering of applications with noise,DBSCAN)对点云进行实例分割,该算法将空间中距离小于设定阈值ε的点归为同一簇,从而将场景点云划分为多个独立的点云团,每个点云团对应一个潜在的物体实例。
如图6所示,为验证所提出的动态语义建图算法的有效性,实验选用TUM数据集中的Structure vs. Texture与Dynamic Objects两个类别进行系统评估。Structure vs. Texture类别包含多种具有不同结构特征与纹理分布的典型室内场景,适用于评估算法在弱纹理、重复结构等挑战性环境中的建图稳定性与特征利用能力。而Dynamic Objects类别则专门针对动态环境设计,序列中包含行人移动等真实动态干扰,能够有效检验算法在动态目标识别、语义地图更新与长期一致性保持方面的表现。通过结合两类场景,该实验既可验证算法在静态复杂环境中的通用性,也能深入考察其在动态场景下的鲁棒性与实用性,从而全面评估所提方法的综合性能。
为了更好地展示轻量化地图(八叉树地图)的建图质量与环境结构信息,采用TUM数据集中的 Structure vs. Texture 类别来进行八叉树建图实验。如图7所示,对收集到的点云进行预处理,包括去除噪声、滤波处理以及降低采样率等步骤,以提高数据质量和后续处理的效率。接着,将稠密点云数据转换为八叉树地图,如图7(b)所示,八叉树地图将空间划分为多个立方体单元,每个单元可以根据其包含的点云信息进行分类,如地面、墙壁等,这有助于提取环境中的语义信息。此外,利用高精度三维扫描设备获取室内3D环境的三维坐标及其对应的强度和颜色信息,图7(c)所示为赋予了强度和颜色信息的RGB八叉树地图。
TaketomiT, UchiyamaH, IkedaS. Visual SLAM algorithms: a survey from 2010 to 2016[J]. IPSJ Transactions on Computer Vision and Applications, 2017, 9(1): 16.
ZhangLei, XuXiaobin, CaoChenfei, et al. Robot pose estimation method based on image and point cloud fusion with dynamic feature elimination[J]. Chinese Journal of Lasers, 2022, 49(6): 0610001.
[6]
HanS, XiZ. Dynamic scene semantics SLAM based on semantic segmentation[J]. IEEE Access, 2020, 8: 43563-43570.
ChenGuanhua, LiBo, ZhuZhengtao. Research on 3D point cloud segmentation algorithm of air conditioner radiator V-slots based on improved PointNet[J]. Science Technology and Engineering, 2024, 24(5): 1963-1971.
ChenXiufeng, ZhaoFengyang, WangChengxin, et al. Car-following model and stability analysis based on dynamic safety distance[J]. Science Technology and Engineering, 2024, 24(5): 2105-2110.
[11]
SaputraM R U, MarkhamA, TrigoniN. Visual SLAM and structure from motion in dynamic environments: a survey[J]. ACM Computing Surveys (CSUR), 2018, 51(2): 1-36.
[12]
AsgariT S, AbhishekK, CohenJ P, et al. Deep semantic segmentation of natural and medical images: a review[J]. Artificial Intelligence Review, 2021, 54(1): 137-178.
[13]
TangJ, FolkessonJ, JensfeltP. Geometric correspondence network for camera motion estimation[J]. IEEE Robotics and Automation Letters, 2018, 3(2): 1010-1017.
TianXuan, WangLiang, DingQi. A survey of image semantic segmentation methods based on deep learning[J]. Journal of Software, 2019, 30(2): 440-468.
[16]
HeK, GkioxariG, DollárP, et al. Mask R-CNN[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 2961-2969.
[17]
BolyaD, ZhouC, XiaoF, et al. Yolact: real-time instance segmentation[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 9157-9166.
[18]
YanL, LiuD, XiangQ, et al. PSP net-based automatic segmentation network model for prostate magnetic resonance imaging[J]. Computer Methods and Programs in Biomedicine, 2021, 207: 106211.
[19]
PengH, XueC, ShaoY, et al. Semantic segmentation of litchi branches using DeepLabV3+ model[J]. IEEE Access, 2020, 8: 164546-164555.
ZhangBo, ZhangMiaohui, ChenYunzhong. Crop pest recognition based on spatial pyramid pooling and deep convolutional neural network[J]. Transactions of the Chinese Society of Agricultural Engineering, 2019, 35(19): 209-215.
[22]
王霞, 左一凡. 视觉 SLAM 研究进展[J]. 智能系统学报, 2020, 15(5): 825-834.
[23]
WangXia, ZuoYifan. Research progress of visual SLAM[J]. CAAI Transactions on Intelligent Systems, 2020, 15(5): 825-834.
[24]
ShafiqueK, LeeM W, HaeringN. A rank constrained continuous formulation of multi-frame multi-target tracking problem[C]// 2008 IEEE Conference on Computer Vision and Pattern Recognition. Anchorage: IEEE, 2008: 1-8.
[25]
ZhangJ, HeneinM, MahonyR, et al. VDO-SLAM: a visual dynamic object-aware SLAM system[J]. arXiv prePrint, arXiv: 2005. 11052, 2020.
[26]
HossenM K, TuliS H. A surveillance system based on motion detection and motion estimation using optical flow[C]// 2016 5th International Conference on Informatics, Electronics and Vision (ICIEV). Dhaka: IEEE, 2016: 646-651.
[27]
MenzeM, HeipkeC, GeigerA. Object scene flow[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2018, 140: 60-76.
ZhangRongfen, YuanWenhao, LuJin, et al. A semantic octree map construction method for visual SLAM in indoor dynamic scenes[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1811003.