动态场景下基于特征点筛选的视觉SLAM算法

doi:10.16182/j.issn1004731x.joss.23-1406

摘要/Abstract

摘要：

针对传统的视觉SLAM算法在动态环境下定位精度与鲁棒性低的问题，提出了一种基于特征点筛选的改进动态SLAM算法。该算法以ORB-SLAM3算法为基本框架，增加了动态区域划分与特征点筛选模块。动态区域划分模块使用改进的RT-DETR目标检测算法检测图像中的动态物体并根据检测框划分动态区域，特征点筛选模块使用对极约束以及光流法筛选掉运动的动态物体身上的特征点，保留静止的动态物体上以及目标检测框内属于背景部分的特征点，参与后续的位姿优化。改进的算法尽可能的保留了更多的有效特征点参与相机位姿优化，实验结果表明：改进的算法在高动态环境下绝对轨迹误差的RMSE值，能够达到平均90%以上提升的同时保持实时运行。

关键词: 光流, 对极约束, 视觉SLAM, 动态环境, RT-DETR, PP-LCNet

Abstract:

To address low positioning accuracy and robustness in traditional visual SLAM algorithms under dynamic conditions, this paper proposes an improved dynamic SLAM algorithm based on feature point selection. Built upon the ORB-SLAM3 framework, it incorporates dynamic region partitioning and feature point filtering.The dynamic region partitioning module utilizes an enhanced RT-DETR object detection algorithm to detect dynamic objects in the images and divides the dynamic regions based on the detection boxes. The feature point selection module utilizes epipolar constraints and optical flow methods to filter out feature points on moving objects, retaining stationary dynamic objects and background points within the detected object bounding boxes for subsequent pose optimization. The improved algorithm strives to retain as many effective feature points as possible for camera pose optimization. Experimental results show that the proposed algorithm improves the RMSE of absolute trajectory error in highly dynamic environments by over 90% on average, while maintaining real-time operation.

Key words: optical flow, epipolar constraint, visual SLAM, dynamic environment, RT-DETR, PP-LCNet

中图分类号:

TP242

姜丽梅,陈信威 . 动态场景下基于特征点筛选的视觉SLAM算法[J]. 系统仿真学报, 2025, 37(3): 753-762.

Jiang Limei,Chen Xinwei . Visual SLAM Algorithm Based on Feature Point Selection in Dynamic Scenes[J]. Journal of System Simulation, 2025, 37(3): 753-762.

图/表 10

图1

图2

图3

表1

图4

表2

表3

表4

表5

表6

参考文献 22

1	Gaia Jeremías, Orosco Eugenio, Rossomando Francisco, et al. Mapping the Landscape of SLAM Research: A Review[J]. IEEE Latin America Transactions, 2023, 21(12): 1313-1336.
2	Campos Carlos, Elvira Richard, J Gómez Rodríguez Juan, et al. ORB-SLAM3: An Accurate Open-source Library for Visual, Visual-inertial, and Multimap SLAM[J]. IEEE Transactions on Robotics, 2021, 37(6): 1874-1890.
3	Qin Tong, Li Peiliang, Shen Shaojie. VINS-mono: A Robust and Versatile Monocular Visual-inertial State Estimator[J]. IEEE Transactions on Robotics, 2018, 34(4): 1004-1020.
4	Saputra M R U, Markham A, Trigoni N. Visual SLAM and Structure from Motion in Dynamic Environments: A Survey[J]. ACM Computing Surveys, 2018, 51(2): 37.
5	赵燕成, 魏天旭, 仝棣, 等. 动态场景下融合YOLOv5s的视觉SLAM算法研究[J]. 无线电工程, 2024, 54(4): 900-910.
	Zhao Yancheng, Wei Tianxu, Tong Di, et al. Research on Visual SLAM Algorithm Incorporating YOLOv5s in Dynamic Scenes[J]. Radio Engineering, 2024, 54(4): 900-910.
6	Pumarola Albert, Vakhitov Alexander, Agudo Antonio, et al. PL-SLAM: Real-time Monocular Visual SLAM with Points and Lines[C]//2017 IEEE International Conference on Robotics and Automation (ICRA). Piscataway: IEEE, 2017: 4503-4508.
7	Su Deng, Chong Dehong. A RGB-D SLAM Algorithm for Indoor Dynamic Scene[EB/OL]. (2020-11-28) [2023-11-17]. .
8	Cheng Jiyu, Sun Yuxiang, Meng Qinghu. Improving Monocular Visual SLAM in Dynamic Environments: An Optical-flow-based Approach[J]. Advanced Robotics, 2019, 33(12): 576-589.
9	Samadzadeh Ali, Nickabadi Ahmad. SRVIO: Super Robust Visual Inertial Odometry for Dynamic Environments and Challenging Loop-closure Conditions[J]. IEEE Transactions on Robotics, 2023, 39(4): 2878-2891.
10	Bescos Berta, Fácil José M, Civera Javier, et al. DynaSLAM: Tracking, Mapping, and Inpainting in Dynamic Scenes[J]. IEEE Robotics and Automation Letters, 2018, 3(4): 4076-4083.
11	Mur-Artal Raúl, Tardós Juan D. ORB-SLAM2: An Open-source SLAM System for Monocular, Stereo, and RGB-D Cameras[J]. IEEE Transactions on Robotics, 2017, 33(5): 1255-1262.
12	Yu Chao, Liu Zuxin, Liu Xinjun, et al. DS-SLAM: A Semantic Visual SLAM Towards Dynamic Environments[C]//2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Piscataway: IEEE, 2018: 1168-1174.
13	Su Peng, Luo Suyun, Huang Xiaoci. Real-time Dynamic SLAM Algorithm Based on Deep Learning[J]. IEEE Access, 2022, 10: 87754-87766.
14	Lowe David G. Distinctive Image Features from Scale-invariant Keypoints[J]. International Journal of Computer Vision, 2004, 60(2): 91-110.
15	Bay Herbert, Tuytelaars Tinne, Luc Van Gool. SURF: Speeded up Robust Features[C]//Computer Vision-ECCV 2006. Berlin: Springer Berlin Heidelberg, 2006: 404-417.
16	Rublee E, Rabaud V, Konolige K, et al. ORB: An Efficient Alternative to SIFT or SURF[C]//2011 International Conference on Computer Vision. Piscataway: IEEE, 2011: 2564-2571.
17	Zhao Yian, Wenyu Lü, Xu Shangliang, et al. DETRs Beat YOLOs on Real-time Object Detection[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 16965-16974.
18	Cui Cheng, Gao Tingquan, Wei Shengyu, et al. PP-LCNet: A Lightweight CPU Convolutional Neural Network[EB/OL]. (2021-09-17) [2023-11-17]. .
19	Bouguet J Y. Pyramidal Implementation of the Lucas Kanade Feature Tracker Description of the Algorithm[EB/OL]. (1999) [2023-11-17]. .
20	Meinhardt-Llopis Enric, Javier Sánchez Pérez, Kondermann Daniel. Horn-schunck Optical Flow with a Multi-scale Strategy[J]. Image Processing on Line, 2013, 3: 151-172.
21	Lin T Y, Maire M, Belongie S, et al. Microsoft COCO: Common Objects in Context[C]//Computer Vision-ECCV 2014. Cham: Springer International Publishing, 2014: 740-755.
22	Sturm Jürgen, Engelhard Nikolas, Endres Felix, et al. A Benchmark for the Evaluation of RGB-D SLAM Systems[C]//2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. Piscataway: IEEE, 2012: 573-580.

模型	平均精度/%	平均精度(重叠度50%)/%	参数量/M	每秒浮点运算次数/G	帧率(RTX4060)/(帧/s)
PPLCNet-RT-DETR	45.0	61.9	10	29	137
RT-DETR-R18	46.5	63.8	20	60	100
RT-DETR-R34	48.9	66.8	31	92	72
RT-DETR-R50	53.1	71.3	42	136	49
RT-DETR-R101	54.3	72.7	76	259	30

序列	ORB-SLAM3		本文算法		Improvement
序列	RMSE	STD	RMSE	STD	RMSE	STD
walking_halfsphere	0.262 9	0.091 3	0.019 6	0.009 7	0.925 4	0.893 7
walking_xyz	0.275 0	0.146 7	0.014 9	0.008 0	0.945 8	0.945 4
sitting_static	0.006 6	0.003 2	0.005 6	0.002 6	0.151 5	0.187 5
sitting_halfsphere	0.022 9	0.012 5	0.015 3	0.008 7	0.331 9	0.304 0

序列	ORB-SLAM3		本文算法		Improvement
序列	RMSE	STD	RMSE	STD	RMSE	STD/%
walking_halfsphere	0.153 4	0.098 4	0.013 5	0.007 8	0.911 9	0.920 7
walking_xyz	0.149 7	0.125 7	0.012 1	0.007 2	0.919 1	0.942 7
sitting_static	0.004 8	0.002 4	0.004 6	0.002 3	0.041 6	0.041 6
sitting_halfsphere	0.008 0	0.004 3	0.007 5	0.004 1	0.062 5	0.046 5

序列	DynaSLAM	DS-SLAM	RTD-SLAM	本文算法
序列	RMSE	RMSE	RMSE	RMSE
walking_halfsphere	0.025 4	0.030 3	-	0.019 6
walking_xyz	0.015 3	0.024 7	0.017 6	0.014 9
walking_static	0.006 8	0.008 1	0.006 9	0.006 9
walking_rpy	0.035 6	0.444 2	0.061 2	0.030 4
sitting_xyz	0.009 4	-	0.008 7	0.008 4
sitting_static	-	0.006 5	0.005 8	0.005 6

算法	网络模型	单帧跟踪时间/ms	GPU
ORB-SLAM3	-	16.4	RTX 4060
DynaSLAM	Mask R-CNN	153.9	RTX 4060
本文算法	PPLCNet-RT-DETR	46.8	RTX 4060
R18-RT-DETR-SLAM	RT-DETR-R18	49.5	RTX 4060